Basic research for the realization of online MEG using SSD

Neurofeedback systems have been found to be effective in the clinical rehabilitation of paralysis. However, most systems exist only for use with EEG, which is cumbersome to apply to patients and has lower spatial resolution than MEG. Furthermore, the best practices for neural data feature extraction and feature selection are not well established. The inclusion of the best performing feature extraction algorithms is critical to the development of clinical neurofeedback systems. Using simultaneously collected MEG and accelerometer data before and during 10 spontaneous finger movements, we performed an in-depth comparison of independent components analysis (ICA) and spatio-spectral decomposition (SSD) algorithms for their individual abilities to isolate movement-relevant features in brain activity. Having restricted raw data to that from sensorimotor rhythm (SMR) frequencies in select MEG sensors over sensorimotor cortex, we compared ICA and SSD components using: (1) 2D topographies, (2) activations over time, (3) and correlations with accelerometer data at both 0ms and 60ms time delays. SSD performed more quickly and produced components that were more highly correlated with the behavioral data than ICA. We will discuss these results and suggestions for application to neurofeedback systems. In particular, we will present detailed visualizations of SSD results and discuss potential strategies and pitfalls for feature selection.