IJCSNS International Journal of Computer Science and Network Security, VOL.10 No.1, January 2010 144 Manuscript received January 5, 2010 Manuscript revised January 20, 2010 A Time-Domain Subspace Technique for Estimating Visual Evoked Potential Latencies Mohd Zuki Yusoff and Nidal Kamel Universiti Teknologi PETRONAS, Bandar Seri Iskandar, 31750 Tronoh, Perak, MALAYSIA Summary Estimating a visual evoked potential (VEP) from the human brain is challenging since its signal-to-noise ratio (SNR) is generally very low. An optimization and eigen-decomposition- based subspace approach has been investigated and tested to estimate the latencies of visual evoked potential (VEP) signals which are highly corrupted by spontaneous electro- encephalogram (EEG) waveforms that can be considered as colored noise. This scheme termed as the generalized subspace approach (GSA) depends on the generalized eigendecomposition of the covariance matrices of the VEP and the colored EEG noise. The subspace algorithm jointly transforms these two correlation matrices into diagonal matrices, which can then be segregated into signal subspace and noise subspace. Enhancement is performed by removing the noise subspace and estimating the clean VEP signal from the remaining signal subspace. Further, GSA has been compared with a third-order correlation (TOC) method, using both realistic simulation and real patient data gathered in a hospital. The simulation results produced by the GSA algorithm show more faithful reproduction of VEP waveforms, and a higher degree of consistencies in detecting the P100, P200, and P300 peaks. Additionally, the results of the real patient data confirm the superiority of GSA over TOC in estimating VEP's P100 latencies, which are used by clinicians to assess the conduction of electrical signals from the subjects' retinas to the visual cortex parts of their brains. Keywords: Visual evoked potentials, signal subspace, time-domain estimator, colored EEG noise, VEP latencies. 1. Introduction A visual evoked potential (VEP) exists when a subject under study is shown a visual stimulation (e.g., a pattern reversal checkerboard). VEP latencies such as the P100’s are used by clinicians to check the integrity of the visual pathways from the retina to the occipital cortex part of the brain [1]. The VEP is not immediately distinguishable from the brain recording because it is buried deep inside the ongoing electroencephalogram (EEG) noise, with a typical signal-to-noise ratio (SNR) of -5 to -10 dB [2, 3, 4, 5, 6]. The post-stimulation EEG which contaminates the VEP is highly colored and its correlation matrix cannot be directly obtained from the observed (corrupted) VEP; the pre-stimulation EEG which exists prior to the application of stimulation is the only sample that can be used to approximate the correlation matrix of the post-stimulation EEG. Conventionally, ensemble averaging (EA) is used to extract the VEPs. For this, hundreds of trials need to be acquired and averaged out to really produce clean VEP estimates; this requires very long recording time that will certainly cause discomfort and fatigue to the subject under study; exhaustion will result in inconsistent formation of VEPs in terms of both amplitudes and latencies. Among the most recent “single-trial” approaches to detect VEPs is a third-order correlation (TOC) technique proposed by Gharieb and Cichocki [7]. This technique performs well in handling white and colored noise whose spectrum does not overlap with that of the desired signal. However, when the signal spectrum overlaps with the highly colored noise spectrum, the efficiency of the TOC-based technique is compromised. The focus of this study is to correctly estimate VEP latencies, instead of VEP amplitudes. In hospitals, doctors are much more concerned about VEP latencies as opposed to VEP amplitudes; the latencies are used by clinicians to assess the visual pathway integrity of the patient under investigation. In general, an approach based on a signal subspace principle performs well in estimating the desired peak positions (i.e., latencies) of a given waveform. The VEP extraction method presented here is inspired by work from a speech enhancement area, originally proposed by Ephraim and Van Trees [8]; the original work dealt primarily with white noise. The incorporation of universal optimization schemes in [8] makes them suitable for our single-trial estimation of VEPs. However, to deal with colored noise, we introduce generalized eigen- decomposition instead of normal eigen-decomposition operation in the underlying signal subspace estimator.