A Multi-Frequency Self-Supervised Fusion Model for EEG-Based Dementia Classification

Brain source localization technology enables precise characterization of the spatial distribution of neural activity, serving as a crucial tool for exploring the pathological mechanisms underlying dementia. However, effectively integrating complementary diagnostic information from source localization features across multiple frequency bands remains a major challenge to enhancing classification performance and model interpretability.

An attention-based multi-frequency self-supervised fusion model (AM-SSF) is proposed to address this issue. Independent contrastive self-supervised encoders are trained for the θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-48 Hz) frequency bands to learn band-specific latent representations. Then, an attention-guided adaptive fusion module is introduced to dynamically allocate band weights through cross-entropy-based supervised optimization, thereby achieving effective cross-band information integration. Finally, a random forest classifier is employed to evaluate the model's performance in distinguishing Alzheimer's disease (AD) from frontotemporal dementia (FTD).

Experimental results show that the proposed framework achieves a classification accuracy of 93.1% under five-fold cross-validation, significantly outperforming baseline methods such as single-band self-supervised learning (SSL) and average pooling fusion. Further analysis of the attention weight distributions revealed that the θ and β bands contributed most to model decision-making, providing interpretability regarding frequency-specific effects.

In summary, the proposed AM-SSF model enhances AD and FTD classification performance while offering valuable insights into the discriminative roles of frequency band features.