Improving methods of diagnosis and prognostication in neurodegenerative diseases through electroencephalography analysis and machine learning

Abstract

Parkinson's disease (PD) is a neurodegenerative disorder for which no definitive diagnostic test exists. Potential biomarkers of PD are widely researched as they might enable earlier diagnosis, personalisation of treatment, and more accurate prognosis for people with PD (PwPD). One potential biomarker is embedded in electroencephalography (EEG). Recently, machine learning (ML) methods have been applied to analysing EEG data to develop diagnostic and prognostic models of PD. The published studies use many methods of pre-processing, ranging from using raw data to advanced data cleaning, featurisation, and classification. This hinders comparing results across studies and, in addition, no one has yet investigated how different methods of pre-processing might influence results. In this PhD thesis in computer science, the aim was to provide insight into the precise effects of using different methods and, ultimately, to give recommendations for collecting, processing, and analysing data, with a particular focus on using EEG to develop diagnostic and prognostic models for PD in the future. Its seven studies investigated methods of data collection, cleaning, outlier removal, featurisation and feature selection, as well as clarifying the effects of personal demographics on classification metrics. Resting state EEG (RSEEG) analysis and ML were addressed in my first five studies. My findings indicated that, ideally, pre-processing should include filtering, electrode removal, and interpolation; that removing outliers and samples containing a large number of outliers improved accuracy but outlier handling in the test set was not necessary; Gradient Boosting Classifier algorithms outperformed Support Vector Machines; and analysis of features extracted from regions of interest (created by grouping electrodes based on location) performed better than analysis of features derived from single electrodes. I also gave evidence of classification outcomes being strongly influenced by sex, age, and disease severity and that performing both feature selection outside of cross-validation and randomising samples when using single epochs from individual participants led to a larger positive bias of classification results when the number of features per sample was smaller. In my sixth study, I then reported on experiments using event-related potentials (ERPs) that classified PwPD and controls in the context of a Go/No-Go task using the Contingent Negative Variation (CNV), showing the difference between early and late stages of the CNV can be used to distinguish PwPD from controls. Finally, a literature review into the use of ERPs to investigate cognitive tasks in people with Multiple Sclerosis (MS) was included to expand conclusions to other neurodegenerative diseases. The findings in this thesis demonstrated the potential for using EEG, and especially RSEEG, as a tool for classifying PwPD and controls, and hence its potential to assist clinicians in managing PwPD. By providing an analysis of the factors that prevent optimal classification, this thesis provided recommendations regarding the most appropriate way to collect and pre-process data, thereby assisting future research investigating potential biomarkers of PD and MS.