An Alzheimer's Disease Diagnostic System Based on Interpretable Graph Neural Networks

Abstract

This invention relates to an Alzheimer's disease diagnostic system based on interpretable graph neural networks, belonging to the field of medical Internet of Things (IoT) technology. It includes a preprocessing module, a local graph construction module, an important feature node selection module, a node selection process display module, a global graph construction module, an Alzheimer's disease diagnosis module, and a global attribution explanation module. It extracts important feature nodes from neuroimaging data based on a self-attention pooling mechanism and elaborates on the local feature selection process. The selected important node information is then integrated with non-imaging data. The Alzheimer's disease diagnosis module is used to predict three categories of Alzheimer's disease. Finally, the global attribution explanation module is used to attribute and explain the global and single-sample features to verify the effectiveness of feature selection. This invention not only improves diagnostic accuracy but also possesses good interpretability, providing clinicians with a more reliable and effective diagnostic aid.

Description

Technical Field

[0001] This invention belongs to the field of medical Internet of Things technology and relates to an Alzheimer's disease diagnostic system based on interpretable graph neural networks. Background Technology

[0002] Alzheimer's disease (AD), an irreversible neurodegenerative disease, is characterized by a persistent decline in memory, cognitive function, and physical abilities. Based on the patient's physical condition and cognitive level, AD progression can be divided into three stages: normal control, mild cognitive impairment, and AD. In clinical diagnosis, the early symptoms of AD and MCI are often subtle, making early identification extremely difficult. MCI, in particular, is a transitional state between normal aging and Alzheimer's disease, with milder cognitive decline that is difficult to distinguish from normal aging. Traditionally, AD diagnosis relies on physicians' clinical experience, cognitive function assessments, and neuroimaging examinations. However, due to significant individual differences, conventional diagnostic methods are insufficient for accurate diagnosis in the early stages of the disease.

[0003] In recent years, many researchers have relied on single-modal data for Alzheimer's disease (AD) prediction, such as magnetic resonance imaging (MRI), structural magnetic resonance imaging (sMRI), and positron emission tomography (PET). While this approach has achieved good results in some cases, single-modal data often cannot fully capture the complex case progression of AD, limiting the predictive accuracy of the models. Therefore, multimodal data has gradually become the focus of AD research, providing more comprehensive case information. Integrating these multimodal data can reveal the complex pathological mechanisms of AD from different perspectives, thereby significantly improving predictive accuracy. However, while multimodal data can improve predictive accuracy, the complex underlying relationships between different types of data make simple feature splicing or fusion methods based on homogeneous features insufficient to capture the deep connections between data. Therefore, effectively fusing multimodal data and constructing highly adaptive models has become crucial for the early diagnosis of Alzheimer's disease.

[0004] Furthermore, while deep learning models, with their numerous parameters and complex hierarchical structures, can achieve outstanding performance across various tasks, these models are often considered "black boxes," making their internal decision-making processes difficult to interpret. This is particularly true in clinical settings involving patient health, where doctors need to understand the model's decision-making rationale to validate and interpret its predictions. Therefore, enhancing the interpretability of models is especially important in the early prediction of Alzheimer's disease (AD). Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a novel multimodal data fusion prediction framework based on local-global interpretable graph neural networks to address the challenges of early AD diagnosis and solve the problems of early AD diagnosis and interpretability.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An Alzheimer's disease diagnostic system based on interpretable graph neural networks includes:

[0008] Preprocessing module: Used to preprocess raw neuroimaging data to obtain standard T1 images;

[0009] Local map construction module: used to segment a standard T1 image into multiple regions of interest to form a local map;

[0010] Important feature node selection module: This module takes the local graph as input and selects important feature nodes through a self-attention-based pooling mechanism.

[0011] Node selection process demonstration module: This module combines the top-k selection strategy in the pooling mechanism with a heatmap to explain the selection process and results of important feature nodes;

[0012] Global graph construction module: used to construct data samples by combining selected important node features with non-imaging features, and then randomly select known and unknown samples from the dataset to construct a global graph;

[0013] Alzheimer's disease diagnosis module: It is used to take the global graph as input to the GNN layer, adaptively learn the relationship between nodes, predict Alzheimer's disease, and obtain the diagnosis result of Alzheimer's disease;

[0014] Global Attribution Explanation Module: Visualizes the impact of features on results through interpretable methods to understand diagnostic decision-making behavior.

[0015] Furthermore, the preprocessing module preprocesses the raw neuroimaging data, specifically including skull removal, cerebellar resection, image segmentation, standardization, and quality monitoring, thereby obtaining standard T1 images.

[0016] Furthermore, the local graph construction module first uses the Harvard-Oxford atlas to segment the preprocessed standard T1 image into multiple regions of interest (ROIs), with each ROI corresponding to a node, to construct a local graph, and simultaneously obtains the feature matrix of each node and the adjacency matrix of the local graph.

[0017] Furthermore, the important feature node selection module performs pooling operations based on the importance of the nodes' topology, that is, it uses a top-k strategy to select nodes and discard unimportant nodes.

[0018] Furthermore, the node selection process display module combines the top-k strategy with a heatmap, using the heatmap to present the nodes after scores are assigned and the selected nodes, thereby demonstrating the node selection process.

[0019] Furthermore, the global graph construction module fuses new imaging node data with non-imaging data to form a sample set of the global graph. It randomly selects α known samples and β unknown samples from the sample set as input S to the graph neural network module and uses the input S to establish relationships in the global graph.

[0020] Furthermore, the Alzheimer's disease diagnosis module takes the global graph as input and calculates the edge probability matrix and edge weight matrix, specifically including:

[0021] Calculate the side probability matrix E using prior knowledge and risk factors:

[0022]

[0023] Where k represents the k-th risk factor feature, K is the number of risk factor features, and e i,j is the i-th row and j-th column of the edge probability matrix E, representing the edge weights between nodes;

[0024] Then, the similarity between different nodes in the node set is calculated using a CNN model, the absolute difference between each feature of two nodes is obtained, and the edge weights are learned from the absolute differences to obtain the edge weight matrix.

[0025]

[0026] in, W represents the edge weight matrix of the l-th layer network. l The weights between nodes in the data;

[0027] This yields a probabilistically constrained self-metric adjacency matrix.

[0028]

[0029] This indicates element-wise multiplication;

[0030] The matrix is ​​normalized using the SOFTMAX layer, and then added to the adjacency operator family. In the middle, node updates are performed; D represents a diagonal matrix, used to preserve the characteristics of the node itself when updating the node;

[0031] Node updates are performed using each operator in the operator family, and the update results are accumulated into G. n (V (l) Then, the accumulated result is compared with the input node V of this layer. (l) Perform connections to obtain the output node V of the GNN layer while preserving the features of the input nodes. (l+1) =[V (l) G n (V (l) )];

[0032] Finally, after changing the feature dimensions in the last layer, the output of the unknown node is directly input into the SOFTMAX layer to normalize the output; the normalized result is the prediction of the unknown node, thus completing the diagnosis and prediction of Alzheimer's disease.

[0033] Furthermore, the global attribution explanation module performs global attribution explanation on the trained model using a model-agnostic local agent interpretable fusion method. Specifically, it uses SHAP to explain the impact of global features on the model output, and employs the LIME method to analyze the impact of each feature on different categories. This explains the impact of each feature on the prediction results and helps to understand the model's diagnostic decision-making behavior.

[0034] The beneficial effects of this invention are as follows: By introducing the Harvard-Oxford atlas to divide brain regions and combining it with a self-attention mechanism to extract key feature nodes from neuroimaging images, it effectively integrates imaging and non-imaging data, achieving comprehensive and accurate AD diagnosis and prediction. In the local processing part, the feature selection process is explained in detail through a Top-k selection strategy and heatmaps. In the global prediction part, SHAP and LIME are used for feature interpretation analysis, revealing the impact of features on the model output and further verifying the effectiveness of feature selection and the influence of each feature on the model's decision-making. Therefore, this invention not only has high prediction accuracy but also good generalization ability and good interpretability.

[0035] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0037] Figure 1 This is an architecture diagram of an Alzheimer's disease diagnostic system based on interpretable graph neural networks;

[0038] Figure 2 This is a flowchart of an Alzheimer's disease diagnostic method based on interpretable graph neural networks. Detailed Implementation

[0039] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0040] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0041] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0042] This invention provides an Alzheimer's disease diagnostic system based on a locally-globally interpretable graph neural network, which improves AD prediction accuracy while increasing interpretability.

[0043] like Figure 1-2 As shown, to reduce the complexity of the prediction system model execution, a self-attention-based important feature node selection module is used to classify important nodes in the imaging images. Then, its output, along with non-imaging data, is simultaneously fed into the Alzheimer's disease diagnosis module. A global node graph is generated and classified using a GNN model to predict the disease state. Furthermore, to make the prediction process more visual and interpretable, a local-global interpreter is used, including a node selection process visualization module and a global attribution interpretation module. The node selection process visualization module is mainly responsible for visualizing the decisions regarding important features, while the global attribution interpretation module explains the entire prediction process to ensure that the decisions made by the method are more reliable and convincing.

[0044] The specific steps of this system are as follows:

[0045] Preprocessing module: Preprocesses the raw neuroimaging data to obtain standard T1 images.

[0046] Due to the unique nature of early AD prediction research, this invention is based on the ADNI database, a neuroimaging technology for Alzheimer's disease. Two thousand samples were randomly selected from the database, including sMRI image data, age, gender, cognitive scores, and other data features, with the aim of utilizing multimodal features to more accurately predict AD. Since the selected sMRI data is raw and contains a large amount of redundant information, preprocessing of the image data was necessary before importing it. The statistical parametric mapping (Spm 12) tool was primarily used to perform skull removal, cerebellar resection, image segmentation, standardization, and quality monitoring to obtain standard T1 images, effectively reducing the time cost of data processing for the model and resulting in more accurate results.

[0047] Local map construction module: Divides the standard T1 image into multiple regions of interest (ROIs) to form a local map.

[0048] To more accurately map imaging data to specific regions, the Harvard-Oxford atlas was used to segment the subject's brain described by the sMRI images into n regions of interest (ROIs), each ROI corresponding to a node, to construct a local map. Correspondingly, the feature matrix of each node was obtained, and the adjacency matrix of the node can be defined as 1 if two nodes are adjacent, and 0 otherwise.

[0049] Important feature node selection module: It takes the local graph as input and selects important feature nodes through a self-attention-based pooling mechanism.

[0050] Pooling is performed based on the node importance and topological structure. This involves using a top-k strategy to select nodes, discarding (1-k)N less important nodes. The pooling ratio is a hyperparameter k∈(0,1], determining the number of nodes to retain, where N is the number of nodes. The adjacency matrix and node features are then updated to obtain new adjacency and feature matrices. The output of the local selection module is then obtained, showing the output features. Where x i Let || be the feature vector of the i-th node, and || represent connection, thus completing the node selection and obtaining the data information of important nodes.

[0051] The node selection process demonstration module combines the top-k selection strategy in the pooling mechanism with a heatmap to explain the selection process and results of important feature nodes.

[0052] The self-attention pooling mechanism used in the important node selection process calculates a weight for each input node, i.e., a self-attention score for each node, and then uses a top-k selection strategy to select the top kN nodes to determine the important nodes. While this mechanism and strategy itself has a certain degree of interpretability, different imaging methods can provide necessary supplementary information to improve the accuracy of disease diagnosis. Therefore, the top-k strategy is combined with a heatmap, using the heatmap to present the nodes after score assignment and the selected nodes, thereby demonstrating the node selection process, enhancing the understanding of the top-k strategy's decision-making behavior, reducing unnecessary data analysis and calculation processes, and providing users with a more transparent and interpretable decision-making process.

[0053] Global graph construction module: Select important node features and non-imaging features to form data samples, and then randomly select known and unknown samples from the dataset to form a global graph.

[0054] The new imaging node data is fused with non-imaging data (age, sex, APOE4 gene information, cognitive test score features) to form sample 'a' of the global graph. The number of nodes in graph G for each sample is N = s + 3 + 1, where s is the number of important points output by the local selection module, 3 represents age, sex, and APOE4 gene information, and 1 represents cognitive test score features. Then, 10 known samples and 1 unknown sample are randomly selected from the sample set as the model input S. c represents the label, C represents the number of categories (early AD involved three categories: NC, MCI, and AD), and the input S is used to establish relationships in the global graph G = {V, E, W}, that is, node v i The initial features are updated, and the initial graph is set to be fully connected with all edge weights of 1.

[0055] Alzheimer's disease diagnosis module: It takes a global graph as input to the model and uses a graph neural network that can adaptively learn the relationships between nodes to predict AD and obtain AD diagnosis results.

[0056] The global graph is input into the GNN layer, and node features are updated through inter-node information transmission to obtain the labels of unknown nodes, thus achieving AD diagnosis and prediction. This mainly includes two steps: calculation of the self-metric connectivity matrix and node updating. Since the input data is multimodal, containing risk factors, MRI data, and label information, the edge probability matrix and edge weight matrix need to be calculated separately before obtaining the probabilistically constrained self-metric connectivity matrix. First, the edge probability matrix E is calculated using prior knowledge and risk factors. Because risk factors have low dimensionality and regular numerical distribution, they are easily influenced by other features; therefore, they can be directly calculated based on prior knowledge. The edge probability matrix calculated using risk factors is then obtained as follows: Where k represents the k-th risk factor feature, K is the number of risk factor features, and e i,j Let be the edge probability matrix E, with row i and column j representing the edge weights between nodes. Then, a CNN model is used to calculate the similarity between different nodes in the node set, obtain the absolute difference between each feature of two nodes, and learn the edge weights from the absolute differences to obtain the edge weight matrix. in, W represents the edge weight matrix of the l-th layer network. l The weights between nodes in the matrix are used to derive the self-metric adjacency matrix with probabilistic constraints. This represents element-wise multiplication, and the matrix is ​​normalized using the SOFTMAX layer. After normalization, it is added to the adjacency operator family. In this context, node updates are performed. D represents a diagonal matrix used to preserve the characteristics of the nodes themselves when updating them.

[0057] Node updates are performed using each operator in the operator family, and the update results are accumulated into G. n (V (l) Then, the accumulated result is compared with the input node V of this layer. (l) Perform connections to obtain the output node V of the GNN layer while preserving the features of the input nodes. (l+1) =[V (l) G n (V (l) Finally, after changing the feature dimensions in the last layer, the output of the unknown nodes is directly input into the SOFTMAX layer for normalization. The normalized result is the prediction of the unknown nodes, thus completing the diagnostic prediction for AD.

[0058] Global Attribution Explanation Module: Visualizes the impact of features on results through interpretable methods, enabling understanding of the model's diagnostic decision-making behavior.

[0059] The trained model is used to perform global attribution explanation through a model-agnostic local proxy interpretable fusion method. Specifically, SHAP is used to explain the impact of global features on the model output. Simultaneously, to further analyze the impact of each feature on each sample, the LIME method is employed to analyze the influence of each feature on different categories. By combining these two methods, the influence of each feature on the prediction results can be explained, and the model's diagnostic decision-making behavior can be understood, thereby better assisting doctors in diagnosis and prediction.

[0060] This invention, based on interpretable artificial intelligence technology and deep learning methods, proposes an interpretable graph neural network diagnostic method for early-stage Alzheimer's disease patients, aiming to improve the accuracy and interpretability of early diagnosis prediction. This method locally selects from imaging data and fuses it with non-imaging data to enhance diagnostic accuracy. Furthermore, to increase interpretability, the local selection module combines an inherently interpretable top-k selection strategy with heatmaps, making the selection process of imaging data nodes transparent. Two interpretable methods are also employed to perform attribution analysis on the model's output, identifying the single-sample features and global features that have the most significant impact on the decision-making outcome, thus assisting doctors in understanding the decision-making process.

[0061] This invention proposes a local-global interpretable graph neural network method that integrates multimodal data for tri-class prediction in early Alzheimer's disease (AD) diagnosis. By introducing the Harvard-Oxford atlas to segment brain regions and combining it with a self-attention mechanism to extract key feature nodes from imaging data, it effectively integrates imaging and non-imaging data, achieving comprehensive and accurate AD prediction. In the local learning part, the feature selection process is explained in detail using a Top-k selection strategy and heatmaps. In the global prediction part, SHAP and LIME are used for feature interpretation analysis, revealing the impact of features on the model output and further validating the effectiveness of feature selection and the influence of each feature on the model's decision. Compared with classic machine learning and deep learning models, this method outperforms them, exhibiting higher generalization ability and better interpretability.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An Alzheimer's disease diagnostic system based on interpretable graph neural networks, characterized in that: include: Preprocessing module: Used to preprocess raw neuroimaging data to obtain standard T1 images; Local map construction module: used to segment a standard T1 image into multiple regions of interest to form a local map; Important feature node selection module: This module takes the local graph as input and selects important feature nodes through a self-attention-based pooling mechanism. Node selection process demonstration module: This module combines the top-k selection strategy in the pooling mechanism with a heatmap to explain the selection process and results of important feature nodes; Global graph construction module: used to construct data samples by combining selected important node features with non-imaging features, and then randomly select known and unknown samples from the dataset to construct a global graph; Alzheimer's disease diagnosis module: It is used to take the global graph as input to the GNN layer, adaptively learn the relationship between nodes, predict Alzheimer's disease, and obtain the diagnosis result of Alzheimer's disease; Global Attribution Explanation Module: Visualizes the impact of features on results through interpretable methods to understand diagnostic decision-making behavior; The local graph construction module first uses the Harvard-Oxford atlas to segment the preprocessed standard T1 image into multiple regions of interest (ROIs), with each ROI corresponding to a node, to construct a local graph. At the same time, it obtains the feature matrix of each node and the adjacency matrix of the local graph. The Alzheimer's disease diagnosis module takes the global graph as input and calculates the edge probability matrix and edge weight matrix, specifically including: Calculate the side probability matrix using prior knowledge and risk factors : in, Indicates the first Individual risk factors characteristics, The number of risk factor characteristics. It is the edge probability matrix The OK Columns represent the edge weights between nodes; Then, the similarity between different nodes in the node set is calculated using a CNN model, the absolute difference between each feature of two nodes is obtained, and the edge weights are learned from the absolute differences to obtain the edge weight matrix. in, Indicates the first Edge weight matrix of layer network The weights between nodes in the data; This yields a probabilistically constrained self-metric adjacency matrix. This indicates element-wise multiplication; The matrix is ​​normalized using the SOFTMAX layer, and then added to the adjacency operator family. In order to perform node updates; This represents a diagonal matrix used to preserve the characteristics of the nodes themselves when updating them. Update nodes using each operator in the operator family and accumulate the update results. Then the accumulated result is compared with the input node of this layer. Perform connections to obtain the output nodes of the GNN layer while preserving the features of the input nodes. ; Finally, after changing the feature dimension in the last layer, the output of the unknown node is directly input into the SOFTMAX layer to normalize the output; the normalized result is the prediction of the unknown node, thus completing the diagnosis and prediction of Alzheimer's disease. The global attribution explanation module performs global attribution explanation on the trained model using a model-agnostic local proxy interpretable fusion method. Specifically, it uses SHAP to explain the impact of global features on the model output, and employs the LIME method to analyze the impact of each feature on different categories. This explains the impact of each feature on the prediction results and helps to understand the model's diagnostic decision-making behavior.

2. The Alzheimer's disease diagnostic system based on interpretable graph neural networks according to claim 1, characterized in that: The preprocessing module preprocesses the raw neuroimaging data, specifically including skull removal, cerebellar resection, image segmentation, standardization, and quality monitoring, thereby obtaining standard T1 images.

3. The Alzheimer's disease diagnostic system based on interpretable graph neural networks according to claim 1, characterized in that: The important feature node selection module performs pooling operations based on the importance of the nodes and the topology, that is, it uses a top-k strategy to select nodes and discard unimportant nodes.

4. The Alzheimer's disease diagnostic system based on interpretable graph neural networks according to claim 1, characterized in that: The node selection process display module combines the top-k strategy with a heatmap, using the heatmap to present the nodes after scores are assigned and the selected nodes, thereby demonstrating the node selection process.

5. The Alzheimer's disease diagnostic system based on interpretable graph neural networks according to claim 1, characterized in that: The global graph construction module fuses new imaging node data with non-imaging data to form a sample set of the global graph. It randomly selects α known samples and β unknown samples from the sample set as input S to the graph neural network module and uses the input S to establish relationships in the global graph.

Images