Electroencephalography (EEG) provides a promising non-invasive approach for Alzheimer's disease (AD) detection. However, cross-subject EEG-based diagnosis remains challenging due to substantial inter-subject variability and the complex pathological characteristics of AD, which involve both dynamic temporal alterations and abnormal brain functional organization. Existing methods often focus on either temporal patterns or spatial connectivity, while the effective interaction between these complementary representations remains insufficiently explored. To address these limitations, we propose STFormer, a Transformer-based spatio-temporal fusion network for subject-independent AD-related EEG classification.

STFormer is designed to jointly model temporal dynamics, brain-region topology, and their interactions through these complementary components. The Multiscale Masked Prediction Transformer (MMPT) captures robust multi-scale temporal representations by incorporating auxiliary masked reconstruction. The Matrix Fusion Graph Convolution (MFGC) learns disease-related brain-network topology by integrating brain-region-aware feature aggregation and complementary connectivity matrices. Furthermore, the Gated Cross-Attention Fusion (GCAF) module performs adaptive bidirectional interaction between temporal and structural representations, enabling more effective spatio-temporal feature integration.

Conducted on three public EEG datasets under subject-independent five-fold cross-validation and compared with eight representative EEG classification methods, STFormer achieves the best overall performance in both binary and three-class classification tasks. Specifically, STFormer improved accuracy and F1-score by 5.67 and 8.39 percentage points in binary classification, and by 12.93 and 15.66 percentage points in three-class classification, respectively.

The results demonstrate that jointly modeling multi-scale temporal dynamics, brain-region topology, and bidirectional spatio-temporal interactions can effectively improve the cross-subject generalization of EEG-based AD detection. STFormer provides a robust and computationally efficient framework for subject-independent EEG analysis and offers potential value for intelligent neurodegenerative disease assessment.