High-density EEG and independent component analysis mixture models distinguish knee contractions from ankle contractions

Abstract

Decoding human motor tasks from single trial electroencephalography (EEG) signals can help scientists better understand cortical neurophysiology and may lead to brain computer interfaces (BCI) for motor augmentation. Spatial characteristics of EEG have been used to distinguish left from right hand motor imagery and motor action. We used independent component analysis (ICA) of EEG to distinguish right knee action from right ankle action. We recorded 264-channel EEG while 5 subjects performed a variety of knee and ankle exercises. An adaptive mixture independent component analysis (ICA) algorithm generated two distinct mixture models from a merged set of EEG signals (including both knee and ankle actions) without prior knowledge of the underlying exercise. The ICA mixture models parsed EEG signals into maximally independent component (IC) processes representing electrocortical sources, muscle sources, and artifacts. We calculated a spatially fixed equivalent current dipole for each IC using an inverse modeling approach. The fit of the models to the single trial EEG signals distinguished knee exercises from ankle exercise with 90% accuracy. For 3 of 5 subjects, accuracy was 100%. Electrocortical current dipole locations revealed significant differences in the knee and ankle mixture models that were consistent with the somatotopy of the tasks. These data demonstrate that EEG mixture models can distinguish motor tasks that have different somatotopic arrangements, even within the same brain hemisphere.