Early Alzheimer’s Disease Detection Using EEG Brain Signals

Abstract

Alzheimer's Disease (AD) is a progressive neurodegenerative disease, and early detection is essential in order for meaningful interventions to improve patient outcomes. While there are diagnostic modalities such as PET scans that can be effective in detecting AD, they tend to be expensive, invasive, or impractical, illustrating the need for more usable alternatives to monitor the progression of AD. Electroencephalography (EEG) is a non-invasive and inexpensive way to monitor brain activity, and it has the potential of being a useful way to identify early markers of AD before significant cognitive decline as well. However, a reliable EEG detection model still must overcome the two significant challenges of a lack of large- scale labeled datasets sufficient for training, and considerable variability in EEG signals across individuals, which can ultimately greatly affect the generalizability of EEG models for helpful detection of AD through these differences. In this paper, we propose a novel deep learning framework to address these challenges with self-supervised learning. Our framework consists of a pre-training phase from a large, unlabeled corpus of general EEG data to extract the universal features of brain signals without needing to depend on scarce AD-specific labels. We then use a fine-tuning phase to adapt the model to the classification problem using a smaller, labeled AD dataset. The model architecture can acquire both spatial and temporal features from EEG signals, which leads to a more comprehensive method of analysis. The experimental results clearly indicate that this approach offers a substantial improvement in AD detecting performance and that there was an increase in F1 scores at both the sample and the subject level over the current state-of-the-art methods. This research outlines the ability of self- supervised contrastive learning to mitigate the problems of data scarcity and inter-subject variability, and can contribute towards more accurate and scalable AD diagnostic tools.