A Nonlinear Manifold Learning & Dynamical System Approach in Characterising Alzheimer’s Disease

Abstract

Alzheimer’s disease, the leading cause of cognitive decline in older individuals, imposes a significant burden on healthcare and the economy. Currently, around 47 million people worldwide are affected by neurocognitive disorders, with a projected triple by 2050. Therefore, a need for cost-effective methods for early diagnosis is vital. AD is adegenerative neurological condition marked by brain disruptions and cognitive decline.Furthermore, the electrophysiology of brain cortical activity is shown to change before physical symptoms occur, suggesting the potential for early detection and intervention.The electroencephalogram (EEG) is a cost-effective and non-invasive technique used to examine the electrophysiology of cortical activity and is extensively studied in relation to neurodegenerative diseases like Alzheimer’s disease. It provides insights into the functioning and integrity of neural circuits indirectly. Additionally, the EEG can detect abnormalities in physiological processes that disrupt brain networks at an early stage,preceding clinical symptoms and visible structural changes in neuroimaging scans like Magnetic Resonance Imaging (MRI).Extensive research has revealed the nonlinear nature of brain activity’s electrophysiology, with complex temporal interactions among brain regions even during rest. Therefore, it is essential to consider spatial and temporal nonlinearities in EEG analysis to assess brain connectivity and dynamic interactions. This thesis examines changes in the brain cortex of individuals with mild to moderate Alzheimer’s disease. Specifically, it aims to contribute to Alzheimer’s disease characterisation through novel applications of nonlinear methods for connectivity and dynamic analysis of cortical interactions using resting-state EEG.Using kernel-based nonlinear manifold learning techniques, Isomap and GPLVM(Isomap-GPLVM), a novel measure of linear and nonlinear connectivity is derived.This measure helps identify significant changes in nonlinear connectivity between specific brain regions in mild to moderate Alzheimer’s disease. Isomap-GPLVM analysis uncovers significant connectivity differences between occipital bipolar channels and other regions (parietal, centro-parietal, and fronto-central) in Alzheimer’s disease and a group of healthy controls. Furthermore, connectivity changes between fronto-parietalEEG channels and the rest of the channels are found to be crucial for Alzheimer’s disease diagnosis. These results align with previous studies using functional MRI (fMRI),resting-state fMRI, and EEG, supporting links to resting-state functional networks in the brain.Using Isomap-GPLVM analysis, significant changes in statistical dependencies between EEG channel pairs are further investigated for directed dynamic nonlinear dependencies using transfer entropy. This novel application of transfer entropy in characterising Alzheimer’s disease with resting-state EEG uncovers increased intra-hemisphericinformation flow between parietal-occipital and centro-parietal-occipital regions, predominantly in the left hemisphere. These findings suggest a potential compensatory mechanism. These findings are consistent with previous studies utilising resting-state EEG and resting-state fMRI.Cross-frequency interactions between different frequency ranges enable the integration of information from various brain regions. Previous studies have identified changes in cross-frequency interactions within the EEG associated with neurodegenerative diseases like Alzheimer’s disease. The findings from using transfer entropy to examine directed nonlinear dependencies (direction of information flow) between important EEG channels are used to learn dynamic nonlinear input-output time-series models. These models are then analysed in the frequency-domain to examine cross-frequency interactions at higher-order nonlinearities. Data-driven modelling and analysis methods from control systems engineering, system identification and frequency response analysis, are used for this purpose

Date of Award	Oct 2023
Original language	English
Awarding Institution	Coventry University
Supervisor	Fei He (Supervisor) & Matthew England (Supervisor)