A Separable Bi-Pyramidal Feature Attention Network to Detect Alzheimer’s Using Electroencephalographic Signals

Abstract

Signal categorization is crucial in many clinical areas, including the diagnosis of Alzheimer’s disease (AD), a common neurological disorder marked by symptoms such as memory loss and speech difficulties. This study focuses on how to distinguish between Alzheimer’s patients and healthy persons using electroencephalogram (EEG) signals, a noninvasive, low-cost diagnostic approach. We describe a novel separable bi-pyramidal feature attentive network (SBPFAN) that extracts multiscale deep attributes from 2-D images of 8-s EEG segments using separable and dilated convolutions (DCs). A feature attention block (FAB) is incorporated at each pyramid level to emphasize notable AD-related characteristics. After concatenating and processing the FAB feature maps through several dense layers, a softmax layer is employed for classification. Two datasets are used in three different experimental setups—subject-dependent, subject-independent, and cross-dataset—to estimate SBPFAN’s performance. Experimental results demonstrate that SBPFAN is effective and holds significant potential for medical and industrial applications in AD detection.