A spatio-spectral analysis of decoding imagined speech from the idle state

Abstract

Studies into speech imagery (SI) classification from electroencephalogram (EEG) data have generally focused on distinguishing imagined words from each other, but accurate discrimination from the idle state, when the user is relaxed, is also necessary for asynchronous brain-computer interfaces (BCIs). In this study, frequency bands and scalp regions most important for distinguishing SI from the idle state were identified and related to underlying neural processes. Power spectral density (PSD) features were extracted from each channel, and a statistical analysis of the features, as well as a classification analysis involving six classifiers, was carried out. The parietal region was identified as the most important scalp region, whilst the delta, theta, and gamma bands were the most important frequency bands. Furthermore, the importance of the alpha band, and of the temporal, frontal-temporal, frontal-central, and parietal regions varied significantly between the SI vs Idle and SI vs SI classification problems, highlighting the importance of including the idle state in SI classification studies.

Clinical Relevance: This study identifies frequency bands and scalp regions that are significantly important for the SI vs Idle classification problem, which is important for asynchronous SI BCIs.