Classification of Evoked Potentials Associated with Error Observation Using Artificial Neural Networks

Abstract

Observation plays an important role in learning processes. Human development takes place through observation. Observational learning studies indicate that the processes through which observation contributes to learning resemble mechanisms contributing to self-action learning. Scalprecorded Evoked Potentials (EPs) reflect brain electrical activity related to processing of stimuli and preparation of responses. An EP waveform is recorded when an incorrect action is committed by a person called Error-Related Negativity (ERN). ERN is also recorded, with a longer latency and reduced amplitude, when errors are not committed but observed by the person whose EPs are recorded. In the present work the performance of a classifier that discriminates between EPs that are produced by observation of correct or incorrect actions is investigated. Initially, first- order statistical features (mean value, standard deviation, kurtosis, skewness, energy, entropy) from the histogram of each EP recording are calculated. Then, the most significant features are selected using the Sequential Floating Forward Selection (SFFS) algorithm. The Artificial Neural Network (ANN) algorithm combined with the leave-one-out technique is used for the classification task. The overall accuracy for the two classes to be differentiated is above 85%. The successful implementation of systems based on the proposed classifier might enable the improvement of the performance of brain-computer interfaces (BCI) that base their action, among other parameters, on the brain signals that the user emits when he/she detects an undesired response of the BCI.