Feature Extraction of Event Related potential based on Time and frequency Domain Analysis

The event related potential is traditionally obtained in time domain by computing ensemble average. However, due to non-stationarity and poor localization of these signals, this may result in erroneous feature extraction. In this present study, a standard database is considered to elucidate this problem. It is shown that a frequency domain decomposition followed by the estimation of spectral distance by measures like ItakuraSaito distance may partially resolve the problem. However, recognizing the contribution of endogenous and exogenous inputs to each event related potential, it is further argued that a Wavelet Packet Decomposition may be more useful since each signal in the frequency domain can be further decomposed into five characteristic domains (delta, theta, alpha, beta, and gamma) and based on the feasibility of contributions from each domain a better feature extraction will be possible.