Distinguishing Concept Categories from Single-Trial Electrophysiological Activity

Differences in the neural representation of conceptual categories (such as buildings, tools, animals, plants, vehicles) are suggested by studies on brain-injury patients and by those using imaging techniques. Evidence that such conceptual distinctions are encoded spatially has been highlighted by fMRI techniques (Haxby et al, 2001), whereas ERP studies have identified differences in the temporal domain (Kiefer, 2001; Paz-Caballero et al., 2006). The study described here uses a machine learning technique (Dalponte et al., 2007), previously applied to Brain-Computer Interaction (BCI) tasks, to show that conceptual categories can be identified from single-trial spectral EEG responses to visually and auditorily presented stimuli in single participants. We found that using features extracted from frequency spectra, the categorial membership of a single stimulus presentation in one of three classes (animals, plants, tools) can be predicted with an average accuracy of 80%.