Graph neural network modeling and causal interpretation method and system for mental disease recognition

Abstract

The invention provides a graph neural network modeling and causal interpretation method for mental disease recognition. The method comprises the following steps: constructing multi-modal data; performing label labeling on the multi-modal data according to the disease type, and forming a data set by labels and the multi-modal data; constructing an initial model, wherein the initial model comprises a graph construction module, a graph neural network module, a causal interpretation module and a prediction and interpretation output module; training the initial model by using the data set to obtain a graph modeling model for disease classification; and inputting the multi-modal data of the patient into the graph modeling model to obtain a prediction result of the patient. The invention also provides a graph neural network modeling and causal interpretation system. The method provided by the invention is suitable for personalized diagnosis and intervention path analysis, and has good practical application potential.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted medicine, and in particular relates to a graph neural network modeling and causal explanation method and system for the identification of mental illnesses. Background Technology

[0002] Early identification and personalized diagnosis of mental illnesses is an important area of ​​research in modern psychiatry, especially in the clinical management of mental illnesses such as depression, schizophrenia, bipolar disorder, and autism spectrum disorder. Constructing an efficient, accurate, and interpretable auxiliary diagnostic system has become a major challenge for both academia and industry.

[0003] With the rapid development of neuroimaging technology, multimodal imaging techniques such as functional magnetic resonance imaging (fMRI), structural magnetic resonance imaging (sMRI), and diffusion tensor imaging (DTI) have been widely used in the study of the mechanisms of mental illnesses and in assisting in their identification. These imaging technologies can not only non-invasively acquire activity patterns during resting or task states, but also provide rich information on brain region structure, functional connectivity, and dynamic interactions, becoming an important data source for constructing biomarkers for mental illnesses.

[0004] Despite this, existing methods for identifying mental illness still face several technical bottlenecks. On the one hand, traditional methods often rely on static whole-brain connectivity matrices or single statistical features of local brain regions, making it difficult to capture the dynamic changes in the brain at different time scales. On the other hand, complex high-order interactions between brain regions are often overlooked, and time-domain, frequency-domain, and potential causal connections have not been effectively modeled. Furthermore, while some studies have introduced deep learning models to improve recognition accuracy, their generally black-box architecture lacks sufficient clinical interpretability and transparency, limiting their widespread deployment in real-world medical scenarios. Simultaneously, multimodal fusion modeling still faces challenges in current practice; existing methods often employ shallow strategies such as simple feature splicing and weighted fusion, making it difficult to deeply model intermodal synergistic relationships. More importantly, existing models typically cannot identify or explain the direct causal impact of different brain region activities on disease states, lacking causal interpretable mechanisms for clinical applications.

[0005] Patent document CN119848768A discloses a depression recognition auxiliary decision-making system based on multimodal fusion, including: a data acquisition module; a data processing module; a feature extraction module; a depression recognition model module; a model training module, wherein the corresponding depression recognition model module is trained; and a model integration module, wherein the model integration module is used to fuse multiple sets of trained depression recognition model modules together to establish a multimodal depression recognition model for depression recognition.

[0006] Patent document CN115721309A discloses a method for early identification of depression based on EEM and P300 electrophysiological indicators. The method for early identification of depression based on EEM and P300 electrophysiological indicators includes the following steps: S1. Collecting and organizing data and constructing a raw dataset, the data being the EWAS analysis results of candidate gene DNA methylation; S2. Preprocessing the raw dataset to obtain an input dataset, dividing the input dataset into a training dataset and a test dataset; S3. Creating an early learning model to identify patients with depression and healthy individuals, and then training the constructed early learning model using the training dataset. Summary of the Invention

[0007] The purpose of this invention is to provide a graph neural network modeling and causal explanation method and system for the identification of mental illnesses. This method can provide more comprehensive and complete prediction results of mental illnesses, so as to provide reliable visual basis for clinical decision support.

[0008] To achieve the first objective of this invention, the following technical solution is provided: a graph neural network modeling and causal explanation method for the identification of mental illnesses, comprising the following steps: Construct multimodal data, including medical images and corresponding clinical assessment indicators and individual information, wherein the medical images include functional magnetic resonance imaging and structural magnetic resonance imaging; Label the multimodal data according to disease type, and combine the labels and multimodal data into a dataset; Construct the initial model, including a graph construction module, a graph neural network module, a causal explanation module, and a prediction and explanation output module; The graph construction module constructs a corresponding brain network graph based on the connectivity relationships of each brain region and the input multimodal data. The graph neural network module includes two parallel main paths and auxiliary paths. The main path is used to extract causal association features corresponding to multi-hop adjacent nodes in the brain network graph, and the auxiliary path is used to extract heterogeneous connection features corresponding to non-adjacent nodes in the brain network graph. The causal explanation module is based on causal association features and heterogeneous connection features, and introduces... Conditional mutual information is used as the objective function to obtain the independent causal contribution of the corresponding node; The prediction and explanation output module outputs prediction results based on the generated causal association features, heterogeneous connection features, and independent causal contributions. The initial model is trained using the dataset to obtain a graph modeling model for disease classification; The patient's multimodal data is input into a graph modeling model to obtain the patient's prediction results.

[0009] This invention constructs a high-dimensional multimodal brain network graph to comprehensively express the structural connections and dynamic functional relationships between brain regions; extracts deep embedded features related to mental illness states through a graph neural network module; further introduces a causal reasoning mechanism based on conditional mutual information to explicitly identify key subgraph structures in the graph structure that are causally related to disease prediction; and finally outputs personalized disease prediction labels and interpretive causal maps, providing reliable visualization basis for clinical decision support.

[0010] Specifically, the process of constructing the brain network map is as follows: Using brain regions as nodes and the adjacency relationships between brain regions as connecting edges, multimodal data are constructed into a time series data matrix and stored in the nodes of the corresponding brain regions.

[0011] Specifically, the nodes are represented in the form of a quadruple graph: ; in, Let E be the set of nodes, where each node represents an anatomical brain region; let E be the set of edges, representing the connections between different brain regions. For a three-dimensional tensor, , This represents multi-channel edge features, which include mutual information, Causality index and frequency domain characteristics; T is the node feature matrix. ,in This represents the length of the time series.

[0012] Specifically, the objective function expression in the causal explanation module is as follows: ; in, Characteristics of causal relationship This is a heterogeneous connectivity feature. Indicates the predicted label (disease category). They are respectively , , The covariance information kernel matrix, Represents the matrix trace operation. express Entropy order.

[0013] Specifically, the auxiliary path employs a graph attention mechanism to dynamically adjust the representation contribution of neighboring nodes, and uses the representation contribution to calculate the heterogeneous connection features corresponding to non-adjacent nodes.

[0014] Specifically, during the training process, a joint loss function is used for end-to-end training. The expression of the joint loss function is as follows: ;in, The classification loss is based on cross-entropy; For causal explanation loss based on conditional mutual information; Regularization terms to control the sparsity of causal subgraphs; To improve the contrast loss for consistent interpretation; These are the weighting coefficients for each loss term.

[0015] To achieve the second objective of this invention, the following technical solution is provided: a graph neural network modeling and causal explanation system for performing the steps of the graph neural network modeling and causal explanation method for mental illness identification described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: We introduce a causal explanation mechanism based on conditional mutual information to distinguish between driving pathological connections and related interference structures, thereby improving the clinical credibility and interpretability of the identification results. By fusing multimodal neuroimaging to construct a high-dimensional brain map, we can characterize the structure-function coupling pattern and uncover potential pathological mechanisms. We output individual-level causal subgraphs and combine them with a joint loss optimization strategy to balance performance and interpretability. This approach is suitable for personalized diagnosis and intervention path analysis and has good potential for practical application. Attached Figure Description

[0017] Figure 1 This embodiment provides a flowchart of a graph neural network modeling and causal explanation method for the identification of mental illness; Figure 2 This is a schematic diagram of the graph construction module provided in this embodiment; Figure 3 This is a schematic diagram of the operation of the graph neural network module provided in this embodiment; Figure 4 This is a schematic diagram of the causal explanation module provided in this embodiment; Figure 5 This is a schematic diagram of the operation of the prediction and interpretation output module provided in this embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown in this embodiment, a graph neural network modeling and causal explanation method for mental illness identification is provided. The specific steps are as follows: Multimodal data was constructed using MRI acquisition, encompassing resting-state functional magnetic resonance imaging (rs-fMRI) for functional connectivity analysis, structural magnetic resonance imaging (sMRI, especially T1-weighted images) for structural atlas construction, supplemented by clinical assessment scales (such as HAMD and PANSS) and demographic information (such as diagnostic category, age, sex, and years of education). Raw data could be sourced from public databases such as LEMON, ABIDE, and HCP, or multicenter clinical platforms, and uniformly converted to standard formats such as NIfTI or BIDS to support cross-platform compatibility and batch processing.

[0020] The initial model is constructed, which includes a graph construction module, a graph neural network module, a causal explanation module, and a prediction and explanation output module.

[0021] The graph construction module is used to construct a brain network graph structure. This structure not only preserves the spatial connectivity between brain regions, but also integrates multidimensional features from the time and frequency domains, providing rich structural representations for subsequent causal reasoning and model training.

[0022] Functional magnetic resonance imaging (fMRI): In this document, it specifically refers to resting-state fMRI (rs-fMRI), which is primarily used for functional connectivity analysis. Structural magnetic resonance imaging (sMRI): In this document, it specifically refers to T1-weighted images, which are mainly used to construct structural maps.

[0023] These two types of imaging data, along with other clinical indicators and individual information, are used in the "Graph Building Module" to construct brain network maps. In short, sMRI provides a "structural map" of the brain, while fMRI provides information on the "functional activity" of different brain regions; together, they constitute the multimodal data required for the model.

[0024] The graph neural network module is used to extract multi-level graph node embedding representations and to uncover potential pathological mechanisms hidden in the brain network by constructing multi-order graph convolutional sub-networks, graph attention mechanisms, and causal separation coding paths. The causal explanation module is used to realize interpretability modeling in the process of mental illness prediction. This module introduces a causal measurement mechanism based on conditional mutual information, calculates the independent influence strength of potential causal contributions on the prediction label through an objective function, thereby effectively stripping away structures that are causally unrelated to the label but have correlations, and improving the causal credibility of the identification results and the explanatory power of the disease mechanism. This "causal contribution" has a very specific purpose in the document and is not merely an intermediate result. The core of this module is the introduction of Rényi Conditional Mutual Information (RCMI) as the objective function. This "causal contribution" (i.e., RCMI) has two main key roles: During training: optimizing the model and removing interference. The "causal contribution" calculated by this module will be used as a loss (i.e., ...). The causal explanation loss based on conditional mutual information is added to the overall "joint loss function". The model is trained using this method. The purpose of this is to "effectively remove structures that are causally unrelated to the labels but correlated" during training, thereby improving the "causal credibility" of the final prediction. At output: an explanation is generated, and the "causal contribution" of the subgraph (i.e., RCMI loss) is extracted. This is directly used by the "Prediction and Interpretation Output Module". As described in Question 2, the system calculates the gradient of this RCMI loss ( This method identifies which brain connections have a significant causal impact on disease prediction and ultimately extracts a "causal subgraph." This extracted "causal subgraph" serves as the "visual basis" or "interpretive causal map" provided by the model, thus achieving the model's interpretability.

[0025] The prediction and explanation output module is used to integrate the prediction classification results and causal explanation information generated by the aforementioned embedded representation, and output multi-level information including disease prediction labels, causal subgraph extraction results, and potential biomarker localization.

[0026] Among them, "cause-subgraph" ( It is not a newly generated graph, but rather extracted from an existing brain network graph (G). Its general generation (extraction) method is as follows: Calculating the gradient response: The system calculates the "Rényi conditional mutual information loss" (…). For "adjacency matrix elements" The gradient response of ), i.e. Obtaining Causal Weights: The magnitude of this gradient response is used to measure the "high causal weight" of the corresponding connection (edge). Filtering and Extraction: The system filters based on this gradient response value, retaining only those connections with significant causal influence (i.e., high causal weight). Subgraph Formation: The set of these filtered "high-weight edges" ultimately constitutes the "causal subgraph". ) The initial model is trained using the dataset to obtain a graph modeling model for disease classification; The patient's multimodal data is input into a graph modeling model to obtain the patient's prediction results.

[0027] More specifically, medical images are typically DICOM format image sequences. To adapt to subsequent image construction and batch modeling workflows, the original images need to undergo standardized format conversion and preprocessing. Specifically, the DICOM format is first converted to NIfTI format, and then the resulting NIfTI format image is input into DPABI for spatial registration and standardization processing, providing a high-quality input foundation for subsequent image composition and modeling.

[0028] The initial model employs a joint loss function to achieve multi-objective joint modeling, as expressed below: ; This mechanism comprehensively considers classification performance, causal interpretability, structural sparsity, and discriminability. During model training, a minibatch strategy is adopted. While maintaining the consistency of sample labels, the model is guided to extract the graph structure pattern with the most causal influence through the separability of the embedding space, and redundant correlation interference is effectively suppressed.

[0029] The loss function guides the embedding space structure through the relationships between samples, as follows: Among them, the causal representation distance Zα of samples with the same label should be minimized; the distance between samples with different labels should be greater than the preset boundary m to improve the class discrimination; I(⋅) represents the indicator function, and m is an adjustable boundary hyperparameter.

[0030] During training, a minibatch strategy is adopted. While maintaining the consistency of sample labels, the model is guided to extract the graph structure patterns with the greatest causal influence by leveraging the separability of the embedding space, and redundant correlation interference is effectively suppressed.

[0031] like Figure 2 As shown, the system first accepts standardized multimodal neuroimaging data as input and converts it into a high-dimensional heterogeneous brain map structure in the graph construction module. The definitions are as follows: ; in, Let E be the set of nodes, where each node represents an anatomical brain region; let E be the set of edges, representing the connections between different brain regions. For a three-dimensional tensor, , This represents multi-channel edge features, which include mutual information, Causality index and frequency domain characteristics; T is the node feature matrix. ,in This represents the length of the time series.

[0032] Each edge between nodes (i,j) carries a C-dimensional multimodal feature vector, defined as follows: ; in Let be the mutual information value between node i and node j. Granger causality indicators are used to characterize the direction of information flow. For a specific frequency band Power spectral density (such as Delta, Theta, Alpha, Beta); This is a node time series matrix, for each node. Associate a BOLD signal time series of length D.

[0033] like Figure 3 As shown, this is the graph neural network module provided in this embodiment. Figure 3 (a) in the diagram represents the framework of the graph neural network module. The multimodal fusion brain map structure output by the graph construction module is used as input. A dual-pathway graph neural coding structure is introduced to model the embedding path that has a causal relationship with the mental illness label and the interference path that only has a correlation but no causal relationship, so as to achieve causal separation of feature representation.

[0034] like Figure 3 As shown in (b), the main path encoder uses a multi-order graph convolutional network (k-hop GCN), as shown in the following formula: ; in For the normalized adjacency matrix, For the first Layer Learnable weight parameters corresponding to the convolution kernel of order 1; This represents the maximum convolution order (i.e., the maximum number of hops). For activation functions (such as ELU); The path is used to extract causal association features in the multi-hop adjacency structure, which is the causal embedding feature of the l-th layer.

[0035] Auxiliary Path (Non-causal Path): This path shares the input graph structure and features with the main path, but uses a parameter-independent encoder to model potential patterns that are not causally related to label Y but are statistically correlated. It shares the graph input structure but has independent parameter settings, focusing on modeling potential statistically relevant information to enhance the comparative ability of causal identification.

[0036] like Figure 3 As shown in (c), the module incorporates a graph attention mechanism (GAT), and the update formula for the GAT layer is as follows: ; Attention weight The calculation is as follows: ; This mechanism adaptively learns the weight allocation among neighboring nodes, thereby improving the model's robustness and discriminative ability under heterogeneous connectivity features.

[0037] like Figure 4 As shown, this is the causal explanation module mentioned in this embodiment, which introduces... Rényi Conditional Mutual Information (RCMI) serves as a core causal metric to measure latent representations. For predicted labels The independent causal contribution of the variable in controlling for disturbances. The metric under the given conditions is expressed as follows: ; in, Characteristics of causal relationship This is a heterogeneous connectivity feature. Indicates the predicted label (disease category). They are respectively , , The covariance information kernel matrix, Represents the matrix trace operation. express Entropy order.

[0038] like Figure 5 As shown, this is the prediction and interpretation output module mentioned in this embodiment, which includes classification prediction branch, causal subgraph extraction, and biomarker localization.

[0039] Among them, the classification prediction branch is based on causal embedding representation. The disease label prediction for an individual is generated by using a linear transformation and a sigmoid activation function to output the predicted probability, in the following form: ; in, is the learnable weight parameter; b is the bias peak; This represents the predicted probability of disease occurrence (e.g., depression vs. health, 0 / 1 label).

[0040] Cause graph extraction is achieved by calculating the gradient response of the adjacency matrix elements to the Rényi conditional mutual information loss: ; To identify connections that have a significant causal impact in prediction tasks.

[0041] Based on the magnitude of the gradient response, select the set of edges with high causal weights and extract the causal subgraph: ; Biomarker localization module through multiple individuals Aggregate analysis was performed to statistically analyze the frequency distribution of each edge in the subgraph, and statistical tests (such as a significance threshold of p < 0.01) were applied to screen stable, high-frequency connection pairs. Finally, a key brain region connectivity network was constructed as a candidate map for potential structural / functional biomarkers.

[0042] This embodiment also provides a graph neural network modeling and causal explanation system, which is used to perform the steps of the graph neural network modeling and causal explanation method for mental illness identification provided in the above embodiment.

[0043] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0044] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0045] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A graph neural network modeling and causal explanation method for the identification of mental illnesses, characterized in that, Includes the following steps: Construct multimodal data, including medical images and corresponding clinical assessment indicators and individual information, wherein the medical images include functional magnetic resonance imaging and structural magnetic resonance imaging; Label the multimodal data according to disease type, and combine the labels and multimodal data into a dataset; Construct the initial model, including a graph construction module, a graph neural network module, a causal explanation module, and a prediction and explanation output module; The graph construction module constructs a corresponding brain network graph based on the connectivity relationships of each brain region and the input multimodal data. The graph neural network module includes two parallel main paths and auxiliary paths. The main path is used to extract causal association features corresponding to multi-hop adjacent nodes in the brain network graph, and the auxiliary path is used to extract heterogeneous connection features corresponding to non-adjacent nodes in the brain network graph. The causal explanation module is based on causal association features and heterogeneous connection features, and introduces... Conditional mutual information is used as the objective function to obtain the independent causal contribution of the corresponding node; The prediction and explanation output module outputs prediction results based on the generated causal association features, heterogeneous connection features, and independent causal contributions. The initial model is trained using the dataset to obtain a graph modeling model for disease classification; The patient's multimodal data is input into a graph modeling model to obtain the patient's prediction results.

2. The graph neural network modeling and causal explanation method for mental illness identification as described in claim 1, characterized in that, The process of constructing the brain network map is as follows: Using brain regions as nodes and the adjacency relationships between brain regions as connecting edges, multimodal data are constructed into a time series data matrix and stored in the nodes of the corresponding brain regions.

3. The graph neural network modeling and causal explanation method for mental illness identification according to claim 2, characterized in that, The nodes are represented in the form of a quadruple graph: ； in: Let E be the set of nodes, where each node represents an anatomical brain region; let E be the set of edges, representing the connections between different brain regions. For a three-dimensional tensor, , This represents multi-channel edge features, which include mutual information, Causality index and frequency domain characteristics; T is the node feature matrix. ,in This represents the length of the time series.

4. The graph neural network modeling and causal explanation method for mental illness identification according to claim 1, characterized in that, The objective function expression in the causal explanation module is as follows: ； in, Characteristics of causal relationship This is a heterogeneous connectivity feature. Indicates the predicted label (disease category). They are respectively , , The covariance information kernel matrix, Represents the matrix trace operation. express Entropy order.

5. The graph neural network modeling and causal explanation method for mental illness identification according to claim 1, characterized in that, The main path is obtained by modeling using a multi-order graph convolutional network.

6. The graph neural network modeling and causal explanation method for mental illness identification according to claim 1, characterized in that, The auxiliary path employs a graph attention mechanism to dynamically adjust the representation contribution of neighboring nodes, and uses the representation contribution to calculate the heterogeneous connectivity features corresponding to non-adjacent nodes.

7. The graph neural network modeling and causal explanation method for mental illness identification according to claim 1, characterized in that, During training, a joint loss function is used for end-to-end training. The expression for the joint loss function is as follows: ;in, The classification loss is based on cross-entropy; For causal explanation loss based on conditional mutual information; Regularization terms to control the sparsity of causal subgraphs; To improve the contrast loss for interpretive consistency; These are the weighting coefficients for each loss term.

8. A graph neural network modeling and causal explanation system, characterized in that, The steps are for performing the graph neural network modeling and causal explanation method for the identification of mental illness as described in any one of claims 1 to 7.

Images