The estimation of event-related single trial EEG activity is notoriously difficult but is of growing interest in various areas of cognitive neuroscience, such as multimodal neuroimaging and EEG-based brain computer interfaces. However, an objective evaluation of different approaches is lacking. The present study therefore compared four frequently-used single-trial data filtering procedures: raw sensor amplitudes, regression-based estimation, bandpass filtering, and independent component analysis (ICA). High-density EEG data were recorded from 20 healthy participants in a face recognition task and were analyzed with a focus on the face-selective N170 single-trial event-related potential. Linear discriminant analysis revealed significantly better single-trial estimation for ICA compared to raw sensor amplitudes, whereas the other two approaches did not improve classification accuracy. Further analyses suggested that ICA enabled extraction of a face-sensitive independent component in each participant, which led to the superior performance in single trial estimation. Additionally, we show that the face-sensitive component does not directly represent activity from a neuronal population exclusively involved in face-processing, but rather the activity of a network involved in general visual processing. We conclude that ICA effectively facilitates the separation of physiological trial-by-trial fluctuations from measurement noise, in particular when the process of interest is reliably reflected in components representing the neural signature of interest.
All Science Journal Classification (ASJC) codes
- Cognitive Neuroscience