Generalized Subspace Approach for Measurements of Visual Evoked Potentials

Estimating a visual evoked potential (VEP) from the human brain is challenging since its signal-to-noise ratio (SNR) is generally very low. Visual evoked potentials are conventionally extracted from the spontaneous brain activity by collecting a series of time-locked electroencephalogram (EEG) epochs and performing ensemble averaging on these samples to improve the SNR. However, this multi-trial averaging contributes to loss of distinctive physiological information which may prove useful for thorough optical pathway conduction assessment, disease diagnosis, and other fields of study such as psychology and pharmaceuticals. As such, a VEP estimation scheme based on a single VEP trial which minimizes the information loss and reduces VEP recording time, is highly desirable. In this thesis, two novel variations of generalized subspace approaches (GSAs) have been proposed to estimate VEP's P100, P200 and P300 latencies from colored EEG noise. The proposed methods decompose and decorrelate the corrupted VEP space into signal and noise subspace; VEP enhancement is achieved by removing the noise subspace and estimating the clean VEPs only from the signal subspace. Since EEG is colored noise, implicit and explicit pre-whitening of the corrupted VEP waveform are performed in the proposed algorithms, to resolve diagonalization problems and achieve full VEP space decorrelation. Furthermore, the computation of a proper subspace dimension vital to the optimum extraction of VEPs has been included in GSAs. With the diagonalization and subspace dimension problems resolved, the proposed GSA techniques ultimately form a comparable VEP latency estimation system. Three subspace-based approaches for single-trial VEP latency estimation proposed by various authors have also been evaluated and compared with GSAs. The results of comprehensively simulated data involving SNR from 0 to -11 dB indicate that the GSA schemes outperform the other three methods. The GSA estimators produce the lowest failure rate and average errors, and their performance is relatively independent of the given SNR in contrary to the other methods. The results of fifty real patient data further confirm that both GSAs are better estimators compared to the other studied techniques. With the favorable performance demonstrated by the outcome of the simulated and real patient data, both GSAs have the potentials to be used not only as biomedical signal estimators from the brain, but also as general purpose estimators in any other fields where SNR values are relatively low.