Echo State Networks for Modeling and Classification of EEG Signals in Mental-Task Brain-Computer Interfaces

Constructing non-invasive Brain-Computer Interfaces (BCI) that are practical for use in assistive technology has proven to be a challenging problem. We assert that classification algorithms that are capable of capturing sophisticated spatiotemporal patterns in Electroencephalography (EEG) signals are necessary in order for BCI to deliver fluid and reliable control. Since Echo State Networks (ESN) have been shown to be exceptional at modeling non-linear time-series, we believe that they are wellsuited for this role. Accordingly, we explore the ability of ESN to model and classify EEG recorded during several mental tasks. ESN are first trained to model EEG by forecasting the signals a single step ahead in time. We then take a generative approach to classification where a separate ESN models sample EEG recorded during each mental task. This yields a number of ESN that can be viewed as experts at modeling EEG associated with each task. Novel EEG data are classified by selecting the label corresponding to the model that produces the lowest forecasting error. An offline analysis was conducted using eight-channel EEG recorded from nine participants with no impairments and five participants with severe motor impairments. These experiments demonstrate that ESN can model EEG well, achieving error rates as low as 3% of the signal range. We also show that ESN can be used to discriminate between various mental tasks, achieving two-task classification accuracies as high as 95% and four-task accuracies as high as 65% at two-second intervals. This work demonstrates that ESN are capable of modeling intricate patterns in EEG and that the proposed classification algorithm is a promising candidate for use in BCI.