Connectome graph attention network (CGAT): A multimodal approach for enhanced classification of Alzheimer's disease

This study proposes a model for Alzheimer's Disease (AD) diagnosis called the Connectome Graph Attention Network (CGAT), designed to improve the classification of AD stages. By integrating multimodal neuroimaging data and clinical diagnosis information, CGAT constructs a connectome graph for each individual, with brain regions as nodes and the connections between them as edges. CGAT enhances classification by employing an attention mechanism that highlights critical brain region connections. A unique feature of the model is its capacity to merge features from diverse imaging modalities effectively, using modality-specific loss weighting to ensure balanced data contribution in the final classification. Developed in PyTorch Geometric (PyG), CGAT was optimized through the NNI AutoML toolkit. The model demonstrated a notable AUROC score of 94.00, outperforming other graph-based models like GCN, GraphSAGE, GGNN, and ChebNet, especially in differentiating EMCI from LMCI. Visualizations using t-SNE revealed CGAT's ability to distinctly separate cognitive states, emphasizing important connections among brain regions. Overall, CGAT efficiently combines multimodal neuroimaging data for improved, multi-class AD stage classification. • High Accuracy in Classifying Alzheimer's Disease (AD): The CGAT model effectively classifies AD across different stages: Cognitively Normal (CN), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and AD. It achieves an AUROC score of 94.00, outperforming traditional graph-based models like Graph Convolutional Networks (GCN), GraphSAGE, and ChebNet. • Integrating Multiple Neuroimaging Modalities: The model combines data from structural MRI (sMRI), Diffusion Tensor Imaging (DTI), resting-state fMRI, and PET scans using a cross-modality attention mechanism. This approach allows CGAT to capture both structural and functional changes in the brain, leading to improved early diagnosis and staging of AD. • Key Advantage of Attention Mechanisms: The model dynamically assigns more weight to critical brain regions and connections, adapting based on the significance of features from different imaging modalities. This helps in detecting subtle neurodegenerative changes, particularly in the EMCI and LMCI stages, which are often difficult to distinguish with unimodal models. • Comparison with Traditional Graph-Based Models: Models like GCN and GraphSAGE struggle with integrating multimodal data, as they lack adaptive attention mechanisms.t-SNE visualizations show that CGAT clearly separates cognitive states (e.g., CN vs. AD), while unimodal models and traditional graph-based approaches show more overlap, especially between EMCI and LMCI. • Future Research Possibilities: Investigating attention weights assigned to different brain regions could help identify key biomarkers for early AD detection. Combining clinical data with neuroimaging could improve the model's interpretability and provide a more comprehensive understanding of disease progression. Future studies could refine CGAT to develop better diagnostic tools and gain deeper insights into neurodegenerative diseases.