Extracting single trial event related potentials

Established techniques for the analysis of event-related potentials (ERPs) involve averaging time-locked sections of the EEG signal over many trials to obtain the ERP waveform. Such techniques are necessary as single-trial ERP signals are generally obscured by ongoing EEG activity, with signal-to-noise ratios below -10 dB. The averaging process does not preserve changes in the ERP signal from trial to trial. As these may provide clinically useful information, several methods have been developed to enable the extraction of single-trial ERPs. A quantitative method was developed for testing the accuracy of single-trial ERP estimates. ERP signals were simulated by using a piecewise model of an ERP known as the contingent negative variation (CNV). By randomly varying parameters of the model, a large number of unique ERPs were generated. Each of these was embedded in simulated EEG noise modelled as an autoregressive (AR) process driven by white noise. The known simulated ERPs were compared to the corresponding estimates produced by single-trial ERP extraction techniques in terms of the amount of distortion introduced. The above method was used to test a number of single-trial ERP extraction techniques. These were time-sequence adaptive filtering, singularity detection using wavelets, adaptive multi-resolution analysis, and a modification of the multi-resolution analysis technique. None of the methods extracted sufficiently accurate waveforms from single-trial ERPs contaminated with realistic EEG noise. Improved, but still unsatisfactory, ERP estimates were obtained when the AR EEG noise was replaced by Gaussian noise