Assessing EEG neuroimaging with machine learning

Abstract

Neuroimaging techniques can give novel insights into the nature of human cognition. We do not wish only to label patterns of activity as potentially associated with a cognitive process, but also to probe this in detail, so as to better examine how it may inform mechanistic theories of cognition. A possible approach towards this goal is to extend EEG 'brain-computer interface' (BCI) tools - where motor movement intent is classified from brain activity - to also investigate visual cognition experiments. We hypothesised that, building on BCI techniques, information from visual object tasks could be classified from EEG data. This could allow novel experimental designs to probe visual information processing in the brain. This can be tested and falsified by application of machine learning algorithms to EEG data from a visual experiment, and quantified by scoring the accuracy at which trials can be correctly classified. Further, we hypothesise that ICA can be used for source-separation of EEG data to produce putative activity patterns associated with visual process mechanisms. Detailed profiling of these ICA sources could be informative to the nature of visual cognition in a way that is not accessible through other means. While ICA has been used previously in removing 'noise' from EEG data, profiling the relation of common ICA sources to cognitive processing appears less well explored. This can be tested and falsified by using ICA sources as training data for the machine learning, and quantified by scoring the accuracy at which trials can be correctly classified using this data, while also comparing this with the equivalent EEG data. We find that machine learning techniques can classify the presence or absence of visual stimuli at 85% accuracy (0.65 AUC) using a single optimised channel of EEG data, and this improves to 87% (0.7 AUC) using data from an equivalent single ICA source. We identify data from this ICA source at time period around 75-125 ms post-stimuli presentation as greatly more informative in decoding the trial label. The most informative ICA source is located in the central occipital region and typically has prominent 10-12Hz synchrony and a -5 μV ERP dip at around 100ms. This appears to be the best predictor of trial identity in our experiment. With these findings, we then explore further experimental designs to investigate ongoing visual attention and perception, attempting online classification of vision using these techniques and IC sources. We discuss how these relate to standard EEG landmarks such as the N170 and P300, and compare their use. With this thesis, we explore this methodology of quantifying EEG neuroimaging data with machine learning separation and classification and discuss how this can be used to investigate visual cognition. We hope the greater information from EEG analyses with predictive power of each ICA source quantified by machine learning separation and classification and discuss how this can be used to investigate visual cognition. We hope the greater information from EEG analyses with predictive power of each ICA source quantified by machine learning might give insight and constraints for macro level models of visual cognition.