Multi-trial extended subspace-based approach for visually evoked potentials (VEPs) extraction

Objectives: The extraction of the VEP signal from the brain background noise a challenging issue because of the low SNR values. The conventional method of ensemble averaging (EA) does improve the SNR, but at the expense of longer recording time. On the other hand the recently proposed single-trial subspace-based technique manages to extract the VEPs using single trial but at relatively high failure rate. In this research, we extend the subspace-based techniques to multi-trails in order to reduce the failure rate of the subspace-based techniques and to approach EA performance with less number of trials.
Method: The EEG data is first averaged with limited number of trials (10~15 trials) in order to enhance the SNR to the neighbourhood of -3 dB. Then the EEG covariance matrix is prewhitened using Cholesky factorization and linear estimation of the clean signal is performed. The subspace of data covariance matrix is then decomposed into signal subspace and noise subspace. Enhancement is performed by nulling the components in the noise subspace and retaining the components in the signal subspace.
Results: The capability of the proposed technique in extracting the clean VEPs, is assessed and compared with ensemble averaging. In the first experiment the comparison is conducted using artificially generated VEP signals corrupted by colored noise. The capabilities of the techniques in detecting the P100, P200, and P300 and estimating their latencies are used to indicate their performances. The proposed technique is run with 12 trials whereas the EA is run with 100 trials. The results show significant improvement to the single-trial subspace-based technique in terms of bias and failure rates and approximately similar behavior to EA. In the second experiment, the two algorithms are used to estimate the latency of P100 for objective evaluation of visual pathway conduction. The proposed technique is run with 12 trials whereas the EA is run with 80 trials. The results indicate close performance by the proposed technique to EA in terms of bias and failure rate.
Conclusion: A multi-trial subspace-based algorithm is proposed to extract the VEPs from the brain background colored noise. The results indicate significant improvement to single-trial subspace-based technique in term of failure rates and close performance to the ensemble averaging with significantly less number of trials.