Foundation Model Improves EEG-Based Alzheimer's Disease Diagnosis

Diagnosing Alzheimer's Disease (AD) is challenging due to its biological heterogeneity, especially for traditional linear methods that struggle with non-linear neural dynamics. This paper introduces a novel diagnostic framework that leverages the Large Brain Model (LaBraM), a powerful foundation model pre-trained on over 2,500 hours of EEG data. The approach integrates LaBraM's high-dimensional latent embeddings with a non-linear Random Forest classifier to effectively identify robust disease markers. Under a rigorous cross-validation protocol, the method achieved an impressive ROC-AUC of 89.36% and a balanced accuracy of 82.44% in differentiating dementia patients from healthy controls, using only 8-second EEG segments. This performance significantly surpasses traditional spectral baselines. Post-hoc analysis confirmed that the model captures clinically validated biomarkers, such as occipital-frontal Alpha and Theta rhythm degradation, and that higher predicted dementia probability correlated with worse cognitive performance and increased clinical severity. These findings underscore the potential of deep latent representations to extract clinically relevant signatures from noisy EEG signals for precise and rapid diagnosis.