Characterising Neurological Disorders with Nonlinear System Identification and Network Analysis

With an increasingly ageing population, neurological disorders (ND), including Alzheimer's and Parkinson's disease (AD and PD), are becoming the second leading cause of death and the world's largest cause of disability-adjusted life years. Currently, incurable ND have a devastating impact on individuals, families and a heavy economic burden on societies. Early diagnosis and longitudinal monitoring of ND, such as for AD, is extremely important for their treatment, care and on-going research. However, current ND diagnosis approaches, such as cognitive and physical assessment, invasive tests (obtaining biological samples), or neuroimaging scans (e.g. positron emission tomography, magnetic resonance imaging), are often either very subjective and uncomfortable, or very capital intensive and time-consuming.

In this project, we propose a new computational framework that integrates novel nonlinear systems engineering and network analysis for the diagnosis and characterisation of ND based on electroencephalography (EEG) recordings. EEG measures brain electrical activity through small electrodes attached to the scalp (with each electrode called an EEG channel). EEG has the advantage of a relatively low cost (i.e. £100's-£10,000's compared to millions of pounds for magnetic resonance imaging), better accessibility and portability, user-friendliness and, importantly, superior temporal resolution (i.e. high sampling rate with millisecond precision).

Current EEG approaches predominantly employ either the analysis of a single EEG channel or the analysis of pairs of channels using simple (linear) methods that cannot capture the full complexity of the information, and focus on a selected local brain region. The novelty of our new approach will be to characterise ND by analysing the brain as a network using non-linear (cross-frequency) methods. Emerging evidence suggests that cross-frequency coupling (CFC), between different frequency bands, is the key mechanism in the integration of (local and global) communication in the brain across spatial-temporal scales, and thus this project seeks to investigate its role in the development and progression of ND.

Our goal will be realised through the deliverables from four technical work packages (WPs), namely: (1) development (for the first time) of a unified framework to identify and quantify CFC from a systems engineering approach (i.e. nonlinear system identification); (2) development of a novel multi-layer cross-frequency network approach and extraction of global network features; (3) identification of important brain regions for nonlinear dynamic analysis, and; (4) the integration of both local nonlinear CFC features and global network features for diagnostic purposes.

Compared with current machine/deep learning techniques (e.g. recurrent or graph neural networks), our proposed novel approach will provide human interpretable results in addition to the standard classification performance metrics. It will uncover whether linear or nonlinear interactions, the type and variation of nonlinear interactions (e.g. CFC, energy transfer) and which brain regions (EEG channels), are involved in neurodegeneration. Such information can be crucial for developing an interpretable, accurate diagnosis and, eventually, the management of ND. For example, knowing the specific CFC and brain regions involved will not only facilitate the diagnosis of PD, but may also help improve the treatment (i.e. deep brain stimulation) through a more accurate stimulation at specific frequency ranges and brain regions.

We will develop the methodology and evaluate the feasibility of our approach based on the analysis of (anonymised) EEG data collected from AD and PD patients and healthy controls, through the close collaboration and guidance from our project partners, including clinical neurologists at NHS Royal Devon and Exeter Hospital and the University of Sheffield.