The amplitude of Event-Related Potentials (ERP) is significantly lower than the amplitude of the background activity and the non-neural electrical noise. This makes it difficult to extract the ERP signal and averaging is typically used. In
this study we used a new algorithm in order to extract single trials without averaging by solving the blind source separation (BSS) problem.
ERP signals can be sparsely represented using the short time Fourier transform (STFT), which is a plausible assumption for physiological signals, considering the nature of their activation. Therefore we use a blind
source separation method based on the sparseness of the STFT coefficients.
A comparison was made between results of separation with the sparse decomposition and with the more widely used algorithm - "Infomax". In addition, two possible improvements were examined: one by using a template signal and the
second by using wavelet decomposition.
The performance of the two algorithms was examined with simulation signals using artificial ERP signals planted in raw EEG as well as with actual EEG signals. The separated sources showed that artifacts such as eye movement, eye
blink, ECG and in some subjects muscle activity are represented by separate sources. Major differences were found in extracting the ERP signal, with the best separation results obtained with the sparsity criterion. Frontal activity
(N1 and P2) and parietal activity (P3 wave) were represented separately. The single-trial derived P3 wave correlated well with behavior, always preceding reaction time. The correlation with targets was larger than that with non-target stimuli.