Neural network classification of event related potentials for the design of a new computer interface

In this work we discuss the classification of single event related potentials (ERP) with a neural network classifier. The purpose is to discriminate the YES and NO answers thought by a computer user in response to words flashed in the computer display, and to assess the feasibility of building a new computer interface which we call the cortical mouse.The performance of the neural network is compared with a classifier using an eigenvector based extractign method. In this study the neural network outperformed the other classifier. For some subjects the classification is above 90%, but the introspected variability is still large. INTRODUCTION Event related potentials (ERPs) carry information about the intention of the subject response to external stimuli. This features is being explored in our laboratory to discriminate between YES and NO answers thought by the subject in response to questions presented in the computer display. Our goal is to have quadriplegics control the cursor motion in a computer screen, and allow them to use the computer for education, leisure activities or to help them interact with the external world more efficiently. We called this new interface the cortical mouse, The problem of such project resides in the difficult task of discriminating YES from NO answers using ERPs. ERPs are faint, short transients in the electroencephalogram (EEG). The normal way to cope with the negative signal-to-noise ratio is to average the ERPs. However in our project this has two drawbacks: One is related to the long time involved to make a decision. The other is the subject habituation to a stimulus that may violate the condition of equal ERP shape required in the averaging. We are therefore exploring the possibilities of making the discrimination based on single ERPs. This preliminary study was designed to show that the computer could distinguish simple YES-NO thoughts by processing the evoked response potentials (ERPs) embedded in the electroencephalogram. This report outlines the experimental design, the results from two different pattern classification schemes (an eigenvector based extraction method and a neural network) and gives directions for further study. This work was partially funded by RADC contract # F30602-88-0-0027, and Florida High Technology Council 1990.