Single-trial ERP Quantification Using Neural Networks

Traditional approaches to quantify components in event-related potentials (ERPs) are based on averaging EEG responses. However, this method ignores the trial-to-trial variability in the component’s latency, resulting in a smeared version of the component and underestimates of its amplitude. Different techniques to quantify ERP components in single trials have therefore been described in literature. In this study, two approaches based on neural networks are proposed and their performance was compared with other techniques using simulated data and two experimental datasets. On the simulated dataset, the neural networks outperformed other techniques for most signal-to-noise ratios and resulted in better estimates of the topography and shape of the ERP component. In the first experimental dataset, the highest correlation values between the estimated latencies of the P300 component and the reaction times were obtained using the neural networks. In the second dataset, the single-trial latency estimation techniques showed an amplitude reduction of the N400 effect with age and ascertained this effect could not be attributed to differences in latency variability. These results illustrate the applicability and the added value of neural networks for the quantification of ERP components in individual trials. A limitation, however, is that simulated data is needed to train the neural networks, which can be difficult when the ERP components to be found are not known a priori. Nevertheless, the neural networks-based approaches offer more information on the variability of the timing of the component and result in better estimates of the shape and topography of ERP components.