A machine learning framework for analysis of resting-state EEG in patients

The neural mechanisms underlying most psychiatric disorders remain unclear. Many researchers have employed neuroimaging to investigate the neural differences associated with the disorders, and electroencephalography (EEG) are often chosen due to being noninvasive, low cost, and relatively easy to implement in the clinic or laboratory. However, most of the identified potential biomarkers have yet to be translated to the clinic and many previous studies were limited due to their focus on finding group-mean differences for specific EEG features, thus whether combinations of multiple EEG features could serve as diagnostic biomarkers remain unknown. This thesis presents a machine learning framework developed for analysis of resting-state EEG for biomarker discovery in patients. We implemented EEG signal processing, source localization, computation of an extensive set of commonly utilized EEG features, and unsupervised and supervised machine learning algorithms for dimensionality reduction, feature selection, clustering and predictive modelling. The framework was applied to data from combat-exposed veterans with post-traumatic stress disorder (PTSD) and adults with autism spectrum disorder (ASD). In the PTSD study, we observed significant group-mean differences in some of the spectral EEG features, and the classifier was able to classify the PTSD group with up to 63% balanced test accuracy. Interestingly, clustering the PTSD group into two distinct subtypes revealed one subtype with functional connectivity relatively similar to the combat-exposed control group without PTSD and another subtype with prominent hyperconnectivity. Our classifiers trained to classify each of the subtypes against the control group did not obtain better performance on the subtype with relatively normal connectivity, but the classification of the subtype with hyperconnectivity improved up to 79% balanced test accuracy. Additionally, many of the connectivity features utilized by the classifier trained to classify the subtype with hyperconnectivity were positively correlated with arousal severity scores, one of the central symptom clusters of PTSD, and the subtype with hyperconnectivity had greater arousal scores compared to not only the control group, but also the other subtype. In the ASD study, we observed that the adults with ASD had EEG activity patterns within the typical range of the non-autistic comparison group, with no significant group mean or group-variance differences for any of the EEG features and the best classifier merely obtained 56% balanced test accuracy. We also identified two ASD subtypes, but were unable to derive a clinically meaningful interpretation of the subtypes. Taken together, the novel framework presented in this thesis has been demonstrated in two clinical EEG datasets and as it can readily be expanded to other datasets and disorders, we hope the framework can serve as a stepping-stone for future studies and may pave the way for better identification of quantifiable biomarkers in resting-state EEG.