Improved Estimation of Evoked Potentials Using an Iterative Independent Component Analysis Procedure GEORGE ZOURIDAKIS and DARSHAN IYER Department of Computer Science University of Houston 501 PGH Hall, Houston, TX 77204-3010 USA Abstract: - We have developed an iterative approach based on Independent Component Analysis (ICA) to obtain improved estimates of auditory evoked potentials (AEPs), which represent the electrical response of the brain to auditory stimuli. AEPs are often contaminated by artifacts which reduce the localization accuracy of the sources underlying the surface recordings. We use ICA to separate the activity of neuronal generators from the contribution of the various noise sources, under the assumption that improved AEP estimates will result in improved accuracy in source localization. Preliminary results with actual recordings from normal subjects demonstrate the feasibility of the proposed procedure. Key-Words: - Independent Component Analysis, EEG, Auditory Evoked Potentials 1 Introduction Functional brain mapping visualizes the relationship between brain structures and their function. Functional maps can be obtained using multichannel neurophysiological recordings [10], such as evoked potentials resulting from sensory stimulation, and high resolution magnetic resonance imaging (MRI). The structure of the sources underlying surface recordings of brain activity is very complex. Indeed, in addition to neuronal activity, these recordings often contain biological noise from eye movements, muscle activity, and cardiac activity, as well as extraneous interference, for example, from power lines. Ideally, brain recordings should contain only the activity of the cortical generators activated by the experimental task under study. However, the generators associated with the interference signals are independent of the neuronal generators; thus, artifactual signals can be removed via ICA. This approach has been previously applied to speech [5] and face recognition [1], image processing [2], and in the analysis of several biosignals, including the electrocardiogram [4], electrogastrogram [6], and electroencephalogram [7]. In this paper we propose a new iterative ICA procedure that extends our previous methodology [9, 10] to extract improved estimates of the ‘true’ AEP from multichannel surface recordings. The performance of the proposed procedure is currently being evaluated with simulated signals and actual recordings from normal subjects. 2 Methods Scalp-recorded brain activity provides a noninvasive and relatively inexpensive measure of information processing in the brain. Recent advances in hardware and computational tools make dense-array electroencephalography a very attractive approach to functional brain mapping, because of the high temporal and spatial resolution it can provide. Response estimation relies on averaging procedures which enhance activity that is only temporally locked to the stimulus [8], and at the same time, they wipe out all the fine details (the dynamics) of the process. To understand the spatiotemporal evolution in the patterns of brain activation, one must analyze responses on a single- trial-basis, an approach that posses a great challenge because of low signal-to-noise ratio conditions. We have developed an alternative approach based on selective averaging of single-trial responses: similar responses are first grouped together using unsupervised clustering and then partial EPs are computed by averaging all trials in the same group [9, 10]. As an extension to our previous work, we propose a new iterative ICA approach to separate the surface recording into neuronal activity and extraneous noise. Single-trial responses were recorded from normal individuals in response to