Deviation-Based Graph Neural Networks with Kolmogorov-Arnold Transformations for Multiclass Alzheimer’s Disease Classification

lzheimer’s disease classification across cognitive stages is challenging due to heterogeneous progression. Unlike traditional image-based methods that determine brain connectivity based solely on spatial proximity, we propose a multimodal framework that constructs patient-specific brain graphs using a co-deviation principle. This approach links brain regions that simultaneously deviate from healthy patterns, allowing for a more accurate representation of the joint dysfunction associated with mild cognitive impairment and Alzheimer’s disease. The model combines MRI, PET, and clinical data using a Kolmogorov–Arnold Network integrated into a Graph Neural Network, enabling flexible, interpretable nonlinear modeling. Additionally, an explainability pipeline offers insights at the modality, regional, and connectivity levels. Experiments conducted on the ADNI dataset show that this approach improves multiclass performance compared to traditional image-based baselines. Ablation studies and qualitative analyses further confirm the effectiveness and clinical relevance of the method. The source code is publicly available at https://github.com/MicheleMinervini06/g-mmeh.