Asynchronous, adaptive BCI using movement imagination training and rest-state inference

The current study introduces an adaptive Bayesian learning scheme which discriminates between left hand movement imagination, right hand movement imagination and idle (i.e. "no-command") state in an EEG Brain Computer Interface. Unlike previous BCI designs using minimal training, the user does not have to continuously imagine a movement in order to control a cursor. Rather, the cursor reacts meaningfully only when a trained movement imagination is produced. The algorithmic approach was to compute Gaussian probability distributions in log-variance of main Common Spatial Patterns for each movement class, infer from these a prior distribution of idle-class, and allow each distribution to adapt during feedback BCI performance. By producing a markedly different but complexity constrained partition of feature space than with LDA classifiers, allowing the classifier to adapt and introducing an intermediary state driven by the classifier output through a dynamic control law, 90 % level classification accuracy was achieved with less than 5 seconds activation time from cued onset.