Attention-guided electroencephalography fusion with fractal and nonlinear dynamics decodes early Alzheimer’s biomarkers

Alzheimer’s disease (AD) presents a critical and growing challenge to healthcare systems due to its neurodegenerative nature and its complex, varied effects on cognitive function. Although electroencephalography (EEG) data combined with deep learning methods have shown promise for early detection, existing approaches often struggle to achieve reliable accuracy. This difficulty arises from the subtle and heterogeneous neural patterns associated with AD progression, as well as the heightened risk of overfitting when working with small or imbalanced datasets. To address these limitations, this study introduces a new diagnostic modeling paradigm: the gated recurrent unit–attentive EEG fusion (GAEF) model, designed to effectively capture both spectral and temporal dependencies in EEG data. The GAEF model incorporates an attention mechanism that enhances the model’s focus on the most relevant patterns for distinguishing AD, frontotemporal dementia (FTD), and cognitive normal (CN) cases. Furthermore, the model leverages advanced features rooted in fractal and nonlinear dynamics, including Higuchi fractal dimension and Lyapunov exponents, alongside spectral entropy and power spectral density across various EEG bands, to improve classification boundaries. Experimental results across multiple standard datasets demonstrate the robust generalization capability of the GAEF model, achieving a 98.3% accuracy across AD, FTD, and CN classes in both training and testing without overfitting. These findings underscore the model’s practical suitability for reliable deployment in real-world clinical scenarios, significantly advancing precision in AD diagnosis, facilitating early detection, and enhancing patient care strategies.