Unsupervised Learning of Brain State Dynamics during Emotion Imagery using High-Density EEG

Abstract

Here we assume that emotional states correspond to functional dynamic states of brain and body, and attempt to characterize the appearance of these states in high-density scalp electroencephalographic (EEG) recordings acquired from 31 participants during 1-2 hour sessions, each including fifteen 3-5 min periods of self-induced emotion imagination using the method of guided imagery. EEG offers an objective and high-resolution measurement of whatever portion of cortical electrical dynamics is resolvable from scalp recordings. Despite preliminary progress in EEG-based emotion decoding using supervised machine learning methods, few studies have applied data-driven, unsupervised decomposition approaches to investigate the underlying EEG dynamics by characterizing brain temporal dynamics during emotional experience. This study applies an unsupervised approach - adaptive mixture independent component analysis (adaptive mixture ICA, AMICA) that learns a set of ICA models each accounted for portions of a given multi-channel EEG recording. We demonstrate that 20-model AMICA decomposition can identify distinct EEG patterns or dynamic states active during each of the fifteen emotion-imagery periods. The transition in EEG patterns revealed the time-courses of brain-state dynamics during emotional imagery. These time-courses varied across emotions: "grief" and "happiness" showed more abrupt transitions while "contentment" was nearly indistinguishable from the preceding rest period. The spatial distributions of independent components (ICs) of the AMICA models showed higher similarity within-subject across emotions than within-emotion across subjects. No significant differences in IC distributions were found between positive and negative emotions. However, significant changes in IC distributions during emotional imagery compared to rest were identified in brain areas such as the left prefrontal cortex, the posterior cingulate cortex, the motor cortex, and the visual cortex. The study demonstrates the feasibility of AMICA in modeling high-density and nonstationary EEG and its utility in providing data-driven insights into brain state dynamics during self-paced emotional experiences, which have been difficult to measure. This approach can advance our understanding of highly dynamical emotional processes and improve the performance of EEG-based emotion decoding for affective computing and human-computer interaction.