Utilizing Shuffle Attention for Subject-Wise Split for EEG-Based Alzheimer's Detection

Electroencephalography (EEG) has emerged as a noninvasive and cost-effective modality for the automated detection of Alzheimer’s disease (AD), yet many deep learning approaches report inflated performance due to subject-level data leakage. In this study, we propose a convolutional neural network supplemented with a shuffle attention mechanism for binary classification of AD versus healthy controls, evaluated under strictly enforced subject-wise validation. Raw EEG recordings acquired in a routine clinical setting were denoised using artifact subspace reconstruction and independent component analysis, segmented into overlapping epochs, and transformed into time-frequency representations using the short-time Fourier transform. The proposed model integrates shuffle attention at the final convolutional stage to enhance discriminative spectro-temporal features while suppressing subject-specific noise. Model performance was assessed using both a subject-wise holdout split (70/15/15) and leave-one-subject-out (LOSO) cross-validation. Under subject-wise holdout validation, the model achieved an accuracy of 90.87%, recall of 89.47%, precision of 88.98%, F1 score of 89.23%, and specificity of 92.36%, while LOSO validation highlighted the impact of inter-subject variability. These results demonstrate that combining attention mechanisms with rigorous validation protocols yields clinically realistic performance estimates and supports the potential of EEG-based deep learning models for Alzheimer’s disease classification.