Novel Deep Learning Techniques for the Detection and Classification of Neurodegenerative Diseases using Resting State Electroencephalography

Abstract

   Novel Deep Learning Techniques for the Detection and Classification of Neurodegenerative Diseases using Resting State Electroencephalography Christopher Almeida Neves Neurodegenerative diseases are debilitating conditions that progressively deteriorate the life quality of those affected. Compared with traditional neuroimaging modalities, such as Magnetic Resonance Imaging, Electroencephalography (EEG) can provide a more cost-effective and accessible alternative to help underprivileged populations obtain an early diagnosis of their condition, which is paramount for effective patient care. Resting-state EEG (rs-EEG), which records signals while a subject is at rest, offers an alternative to the commonly used task-based experiments for easier-to-adopt data acquisition protocols. While deep learning techniques have been shown to be effective for automatically classifying most EEG signals, they struggle with modeling the longrange temporal dependencies, complex spatial relationships, and the lack of time-locked events in rs-EEG. Aiming to address these issues, we first propose an explainable Graph Neural Network technique for rs-EEG-based Parkinson’s disease detection. Our method uses structured global convolutions to model long-range dependencies and novel multi-head graph structure learning to capture the complex spatial relationships in EEG data. We also propose a head-wise gradient-weighted graph attention explainer to obtain rich connectivity insights. Our second major contribution leverages recent innovations in state space modeling techniques to classify individuals with dementia, and we explore spectral and spatial approaches for learning relationships between EEG channels for the designated task. Additionally, we probe our model’s outputs with explainability techniques and demonstrate that our model learns physiologically relevant features. This thesis puts forth novel deep-learning methods that show promise in addressing challenges in neurodegenerative disease classification using rs-EEG.