Deep Learning-enhanced EEG Preprocessing for Early Alzheimer’s Disease Detection
Emma Russo1, Michelle Lin2, Carissa Chen1, Irene Harmsen3, Nardin Samuel1
1University of Toronto, 2Cove Neurosciences Inc., 3University of Alberta
Objective:
The objective of this study is to apply deep learning methodology to enhance our automated EEG preprocessing pipeline by improving artifact removal methods, with the goal of enhancing signal quality and classification accuracy for early detection of Alzheimer’s disease.
Background:
The diagnosis and management of Alzheimer’s Disease (AD) represents is one of the most pressing neurological challenges worldwide. Recent studies have applied machine learning to EEG-based classification of Alzheimer’s disease, yet many rely on conventional preprocessing techniques that inadequately address artifact contamination and variability across datasets. Deep learning approaches have demonstrated exceptional performance in image and signal denoising tasks, suggesting their potential to overcome these longstanding limitations in EEG analysis.
Design/Methods:
Conventional preprocessing was compared with an updated pipeline incorporating a deep learning (DL) artifact removal module. The DL component was trained to identify and suppress ocular, myogenic, and motion artifacts directly from raw EEG, enhancing signal-to-noise ratio prior to feature extraction. Feature embeddings were generated using uniform manifold approximation and projection (UMAP), and classification of Alzheimer’s disease (AD) versus healthy controls (HC) was performed using random forest models with nested cross-validation. Performance between pipelines was evaluated by comparing area under the receiver operating characteristic curve (AUROC) across independent datasets.
Results:
Integration of the DL artifact removal module improved classification performance across datasets, yielding higher AUROC values (0.86) relative to conventional preprocessing. Qualitative inspection confirmed enhanced signal fidelity and attenuation of ocular and myogenic artifacts. UMAP embeddings demonstrated greater cluster separation between AD and HC following DL-based preprocessing, consistent with improved discriminative feature representation.
Conclusions:
The findings from this study demonstrate that incorporating deep learning into EEG preprocessing improves artifact suppression, signal quality, and downstream classification of Alzheimer’s disease. These findings support DL-driven denoising as a scalable strategy to enhance reliability and translational utility of EEG biomarkers for early neurodegenerative disease detection.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.