Improving EEG-BCI analysis for low certainty subjects by using dictionary learning

The diagnosis of patients with Disorders Of Consciousness represents a challenge in the clinical routine. Recently, Brain Computer Interfaces based in Electroencephalography (EEG-BCI) have been used to detect signs of consciousness in these patients. This approach allows to discover brain responses to command following. Nevertheless, a reliable BCI strategy must to be able to determine the commands with high levels of certainty. Current results reported in the literature evidence that about 25% of the subjects in which BCI is used may have low performances, near to the chance level, even in collaborating subjects. In this work, we propose a novel approach based on dictionary learning representations aimed to improve performance in low certainty subjects. We propose to introduce an intermediate representation scheme, based on sparse dictionaries, before feature selection step. Our main assumption is that by using these representations we can capture more efficiently the EEG signal structure for subjects responses. The results show that using the new representation the weighted average performance in command following outcome the previews proposed methods of 64.8% to 67.9%. Higher improvements in performance were obtained for low certainly subjects. Our results suggests that this approach may improve BCI-EEG performance in low certainly subjects.