Unveiling Cognitive Decline and Sleep: Deep Learning for Alzheimer's Detection and Latent Representations' Exploration for Pattern-Specific Discovery

The escalating prevalence of Alzheimer's disease (AD) among aging populations underscores the critical need for early detection and intervention to mitigate its effects. Sleep disturbances, particularly prevalent in individuals with mild cognitive impairment (MCI) and AD, suggest a significant link between sleep quality and neurodegenerative diseases. Indeed, neuroscientific evidence emphasizes abnormal brain rhythms during sleep and disruptions in specific brain chemicals as fundamental components of cognitive decline and the onset of AD. Importantly, these abnormalities may manifest before noticeable symptoms, indicating their potential as early indicators of neurodegeneration. Electroencephalogram (EEG) recording during sleep is currently a widely adopted technique for extracting such information. This thesis investigates the relationship between sleep quality and AD using EEG data collected from Swiss populations, including healthy individuals, those with AD, and those with MCI. These EEG data signals were collected during sleep sessions, providing crucial insights into the brain activity of participants during rest. The primary objective is to analyze EEG signals and spectrograms to identify patterns indicative of cognitive decline. The contributions of this thesis are outlined as follows: - Exploring Patterns and Insights of EEG spectrograms: Experimental investigation of various data preprocessing techniques that leverage different sleep stages in the signal's spectrograms to identify potentially significant sleep stages. Additionally, three different neural architectures, Convolutional Neural Network, ResNet, and FocalNet, are considered for the task of classification. - Understanding of latent representations of signals using Dynamical Variational Autoencoders (DVAE): DVAEs are particularly useful for understanding temporal information and temporal relationships in the time series, producing latent representations that could be significantly beneficial for study and analysis. - Analyzing the Potential of Latent Representations in Time Series Classification: Utilizing various classification techniques on diverse latent dimensions with convolutional neural networks such as ResNet to investigate the primary characteristics of latent representations. Conducted at the Decision Neuroscience Lab of the Department of Health Sciences and Technology (D-HEST) of the Swiss Federal Institute of Technology (ETH) Zurich, this thesis advances our understanding of AD pathology and pattern-specific discovery, laying the groundwork for early detection strategies.