You are here

Modeling and Classification of EEG using Recurrent Artificial Neural Networks

Here is a brief abstract from the 2012 Front Range Neuroscience Group Annual Meeting summarizing this project.  More details to come soon!


A brain-computer interface (BCI) is a device that enables a user to communicate with a computer system by voluntarily altering their mental state. In one approach for constructing BCI, a user issues instructions to a computer by performing one of several mental tasks. Electroencephalography (EEG) can then be used to monitor brain activity while a machine learning algorithm finds patterns in the EEG that are unique to the mental task the user is performing. Using such a BCI in real time, a person might silently sing a song to move a mouse cursor to the left or perform a mathematical task to move the cursor to the right.

We investigate the use of a novel algorithm for identifying which mental task a user is performing. This algorithm utilizes a type of artificial neural network known as an Echo State Network (ESN). First, ESN are trained to model EEG by forecasting the signal a single step ahead in time. We are able to demonstrate that ESN can produce signals similar to true EEG. Next, we train a separate ESN to model EEG produced while the subject performs each of several mental tasks. In this way, we have a separate ESN that is an expert at forecasting EEG from each task. Previously unseen EEG is labeled by applying each ESN and selecting the label associated with the model that produced the lowest forecasting error.

Data was recorded from 14 subjects using a portable EEG system. Five subjects had severe motor impairments and recording took place in their homes. Nine subjects had no disabilities and recording took place in a laboratory environment. Each subject performed four mental tasks following a visual queue on a computer screen. Five ten-second trials were recorded from each subject for offline analysis. Classification accuracies as high as 65% correct when using all four tasks and as high as 95% for a two-task problem were achieved at two-second intervals. Although the users with disabilities did not perform as well as the users without disabilities, the difference is not statistically significant. Information transfer rates as high as 21 bits per minute were achieved, comparable with state-of-the-art BCI systems.