Differentiating Alzheimer’s disease (AD) from frontotemporal dementia (FTD) using resting-state electroencephalography (EEG) remains clinically challenging because of their overlapping electrophysiological characteristics. Although EEG suits large-scale dementia screening, current method often overestimates performance because of epoch-level data leakage and multiclass feature competition in unified models. We propose a task-decoupled, two-stage hierarchical deep learning framework utilizing multiband EEG microstate dynamics. Continuous microstate sequences, modeled via Hungarian matching to preserve fine-grained temporal information, are processed using a normalizer-free 1D convolutional neural network (1D-CNN-NFNet) integrated with multi-head attention. By decoupling the workflow, Stage 1 performs generalized dementia screening using alpha and delta microstates, achieving an area under the curve (AUC) of 0.851. Stage 2 disentangles AD from FTD using delta and theta dynamics, yielding an AD-locking specificity of 86.1%. Evaluated under a strict subject-level leave-one-subject-out (LOSO) cross-validation protocol, the two-stage framework achieved 63.9% balanced accuracy, outperforming the single-stage baseline (55.4%) with a negligible inference latency of 0.733 ms. Furthermore, attention-based interpretability analysis links frequency-specific microstate alterations to underlying cortical disconnection syndromes. These results demonstrate that the framework provides a reproducible and interpretable auxiliary reference for dementia screening and subtyping in clinical neurology.
Jiang et al. (2026) studied this question.