Single-trial classification of EEG in a visual object task using ICA and machine learning.

Presenting different visual object stimuli can elicit detectable changes in EEG recordings, but this is typically observed only after averaging together data from many trials and many participants. We report results from a simple visual object recognition experiment where independent component analysis (ICA) data processing and machine learning classification were able to correctly distinguish presence of visual stimuli at around 87% (0.70 AUC, p<0.0001) accuracy within single trials, using data from single ICs. Seven subjects observed a series of everyday visual object stimuli while EEG was recorded. The task was to indicate whether or not they recognised each object as familiar to them. EEG or IC data from a subset of initial object presentations was used to train support vector machine (SVM) classifiers, which then generated a label for subsequent data. Task-label classifier accuracy gives a proxy measure of task-related information present in the data used to train. This allows comparison of EEG data processing techniques - here, we found selected single ICs that give higher performance than when classifying from any single scalp EEG channel (0.70 AUC vs 0.65 AUC, p<0.0001). Most of these single selected ICs were found in occipital regions. Scoring a sliding analysis window moving through the time-points of the trial revealed that peak accuracy is when using data from +75 to +125 ms relative to the object appearing on screen. We discuss the use of such classification and potential cognitive implications of differential accuracy on IC activations.