Unsupervised robust detection of behavioral correlates in ECoG

Electrocorticogram (ECoG) based Brain-Computer Interfaces (BCI) provide finer spatial resolution and improved signal-to-noise ratio than its noninvasive counterpart, Electroencephalogram (EEG). This remarkable feature allows for processing in higher spectral bands in order to elucidate more spatially localized encoding mechanisms. We propose an automatic, fully data-driven method to extract relevant neuromodulation events from single-channel, single-trial traces. In particular, our scheme involves two alternating optimizations that resemble k-means; moreover, correntropy is utilized to provide robust estimation and protection against outliers. In this way, we find distinct behavioral correlates in the low-gamma band (76 - 100 Hz) that encode finger flexion movements in a cued task. The results show that correntropy should be used when working with neuronal oscillations due to the high probability of outliers.