EEG-based classification models reveal differential neural processing of words and images

Machine learning methods employing neuroimaging data are useful for monitoring the activation of neural representations. Specifically, they can be used to discern the brain networks engaged in processing specific categories of items. This approach has been used predominantly with functional magnetic resonance imaging data, and more rarely with electroencephalography (EEG) data. Here, we present a task, an analytical pipeline, and a stimulus dataset for investigating category representations using EEG. Participants (N = 30) viewed a series of images and words of objects belonging to five categories (Animals, Tools, Food, Scenes, and Vehicles) and responded when items from the same category were presented consecutively. We trained support vector machines on EEG data within participants and found that both image trials and word trials yielded significant category classification accuracy, with image trials achieving higher accuracy than word trials. When comparing categories in a pair-wise fashion, all pairs were statistically distinguishable for image trials, whereas only one pair was distinguishable for word trials. Parietal and Left Temporal electrodes contributed more to image classification than Frontal and Right Temporal electrodes. Category-specific activity patterns also generalized across participants for image trials. Our data and analytic pipeline yielded high classification accuracies, primarily for image trials, providing support for the utility of EEG data for neural decoding. These methods can be instrumental for exploring the activation and reactivation of neural representations at the category level during wakefulness and, potentially, during offline states.