Estimation of Event-Related Potentials from Single-Trial EEG

Event-related potentials (ERPs) hidden in electroencephalogram (EEG) signals are usually estimated by a superposition and average algorithm, this algorithm makes strong assumptions on the characteristics of ERPs which are not valid. Therefore, it is desirable to discover better methods by which ERPs are estimated from single-trial EEG. In this paper, we proposed a neural network based method to estimate single-trial ERPs. Instead of recovering the ERPs, this method models the amplitudes and latencies of ERPs components. Therefore, a linear generative EEG model was used, which contains a template of ERPs local descriptors (amplitude and latency). To achieve a high accuracy, the classification model was trained by neural networks consist of 1 or 2 hidden layers. We also modified an optimization model defined in the SingleTrialEM algorithm to estimate single-trial ERPs. Then, solving the optimization problem derived from the previous classification model gave estimation of ERPs component descriptors on a single-trial basis. This method was evaluated in simulations. We compared our method with the Woody filter and a SingleTrialEM algorithm with the same simulated data. Results showed that our method is the most effective and capable of extracting ERPs in negative SNR conditions.