A new graph-transformer framework for EEG-based differentiation of Alzheimer's disease and frontotemporal dementia.

Differentiating between Alzheimer's disease (AD), frontotemporal dementia (FTD), and cognitively normal (CN) subjects remains a significant challenge in clinical neurodiagnosis. This study introduces an automated framework that combines electroencephalography (EEG) signal processing with graph-based deep learning (DL) to improve disease classification. The process begins with artifact suppression and a DL-driven filtering model to enhance EEG signal quality. Once filtered, the signals are segmented, and essential features are extracted to build graph representations that reflect brain connectivity patterns. These graphs are then analyzed utilizing a transformer-based graph neural network, enabling accurate classification of AD, FTD, and CN subjects. Results show that the model achieved highly competitive and well-balanced performance in both binary (AD-CN and FTD-CN) and ternary (AD-CN-FTD) classification tasks, with higher accuracy than existing EEG-based diagnostic methods, demonstrating the benefits of integrating signal filtration, graph representations, and transformer architectures. Overall, the findings suggest that this framework can serve as a reliable tool to support clinical decision-making for the early detection and differentiation of neurodegenerative disorders.