A CNN-transformer fusion for EEG-based discrimination of Alzheimer's and frontotemporal dementia

Dementia is a progressive neurodegenerative disorder that severely impacts cognitive functions and daily living, especially in aging populations. Among its subtypes, Alzheimer's disease (AD) and frontotemporal dementia (FTD) exhibit overlapping clinical symptoms, making early and accurate differentiation a critical challenge. Electroencephalography (EEG), as a non-invasive and cost-effective modality, provides valuable insights into the neurophysiological disruptions associated with these conditions. This study aims to develop a robust EEG-based diagnostic framework capable of accurately classifying AD, FTD, and healthy controls (HC) by integrating domain-specific signal processing with advanced deep learning techniques. This study employed a publicly accessible dataset consisting of resting-state EEG recordings from a total of 88 participants, comprising 29 individuals with AD, 23 diagnosed with FTD, and 36 age-matched HC. The proposed model integrates Common Spatial Pattern (CSP) filtering with a sequential modified hybrid architecture that combines Convolutional Neural Networks (CNNs) and a Vision Transformer (ViT). By fusing domain-informed spatial filtering with deep hierarchical feature learning, the model captures both local signal characteristics and global contextual dependencies. A 10-fold cross-validation approach was employed to assess model performance and generalizability. The proposed model achieved notable classification accuracies of 95.86%, 94.76%, 94%, and 92.14% for the AD/HC, FTD/HC, AD/FTD, and AD/FTD/HC classification tasks, respectively. These results underscore the diagnostic potential of EEG-based deep learning frameworks in distinguishing among neurodegenerative conditions and highlight their promise in supporting more precise and individualized clinical interventions. This study presents a novel end-to-end EEG classification pipeline that fuses domain-guided spatial filtering with deep neural feature learning. The promising results suggest that the proposed method could serve as a valuable component in future clinical decision support systems for dementia, contingent upon further validation in real-world clinical settings.