NeuroFormer: A Deep Learning Framework for Alzheimer's Detection Using EEG Signals.

Abstract

Alzheimer's disease (AD), is a prevalent neurodegenerative disorder, characterized by cognitive decline. Alongside AD, and Frontotemporal dementia (FTD) poses significant challenges in clinical diagnosis due to overlapping symptoms and heterogeneous pathological mechanisms. This overlap often complicates timely and accurate diagnosis, which is crucial for effective intervention and treatment planning. To address this, a novel deep learning architecture, NeuroFormer is proposed for the classification of EEG signals captured through non-invasive neurophysiological measurement instruments into three categories: AD, FTD, and healthy controls (CN). The model integrates spectral processing, neural oscillatory dynamics, hierarchical capsule routing, adaptive normalization, and attention-based feature fusion to effectively extract discriminative spatiotemporal features from raw EEG signals. Advanced mechanisms such as dynamic spectral gating and graph convolutions capture both local and global brain activity patterns, while a multi-scale adaptive feature pyramid enhances feature representation. A lightweight classifier decodes the learned features, achieving a classification accuracy of 95.76%, significantly outperforming benchmark models across all key evaluation metrics. These findings demonstrate the strong potential of NeuroFormer as a measurement-driven diagnostic framework for early and accurate classification of dementia-related disorders using EEG instrumentation.