Network Inference And Graph Learning in Characterising Alzheimer’s Disease

Abstract

Alzheimer’s disease (AD), the predominant cause of dementia, presents a growing global challenge as the number of patients continues to rise. This neurodegenerative condition is characterised by disruptions in brain function and a progressive decline incognitive abilities. Research shows that it can be identified before the onset of symptoms, providing an opportunity for early intervention. However, current diagnostic approaches are expensive and inaccessible, highlighting the necessity for a cheap and fast alternative.Electroencephalography (EEG) might offer a viable solution to this need due to its affordability and portability. Although EEG has lower spatial resolution than functional magnetic resonance imaging (fMRI), its excellent temporal resolution in milliseconds makes it a suitable candidate for early detection.This thesis aims to contribute to the characterisation of AD from a graph perspective. This thesis uses graph theory to model EEG signals as graphs, capturing complex interactions between brain regions through functional connectivity (FC) to identify disruptions in brain networks and generate explainable predictions.The first contribution introduces cross-bispectrum, a higher-order spectral analysis method, to reconstruct EEG-based FC graphs. This method can quantify both withinfrequency coupling and cross-frequency coupling. A novel multi layer graph modelling approach integrates information from both coupling types. The analysis reveals significant cross-frequency differences, particularly increased FC in AD cases’ δ-θ coupling.Graph-theoretic analysis proves crucial in understanding the structure and function of cross-frequency brain networks, with vulnerability analysis providing insights into integration and segregation properties.The second contribution explores various FC measures for creating graph-based biomarkers for AD diagnosis using EEG. Graph neural networks (GNNs) are employed to compare the performance of eight FC measures. The study demonstrates that GNN models outperform baseline models and that using FC measures to estimate brain graphs improves the overall performance of GNN. However, no single FC measure consistently outperforms others, highlighting the importance of considering multiple measures for a comprehensive analysis.The third contribution introduces a novel Adaptive gated graph convolutional network (AGGCN), providing explainable predictions for AD diagnosis. AGGCN combines convolution-based node feature enhancement with a correlation-based power spectral density similarity measure. The gated graph convolution dynamically weighs the contribution of various spatial scales, enhancing the model’s interpretability. The proposed AGGCN achieves high accuracy under different conditions and generates consistent explanations of its predictions, offering valuable insights into AD-related alterations of brain networks.In conclusion, this thesis advances our understanding of AD by leveraging EEG and graph theory. The novel contributions, including cross-bispectrum analysis, exploration of various FC measures, and the development of AGGCN, collectively enhance our ability to characterise and diagnose AD from a graph perspective.

Date of Award	May 2024
Original language	English
Awarding Institution	Coventry University
Supervisor	Fei He (Supervisor)