Estimation of Postaverage SNR from Evoked Responses Under Nonstationary Noise

Abstract

In any measure of event-related potentials, it is important to be able to estimate the postaverage signal-to-noise ratio (SNR) in order to assess the quality of the measured signals. The estimated postaverage SNR can be an important detection criteria (as in infant hearing-screening of evoked auditory potentials) and a control factor when comparing signals obtained during different conditions (accounting for residual noise variability). Standard SNR estimation methods, such as the fixed-single-point (Fsp) statistic (C. Elberling and M. Don, "Quality estimation of averaged auditory brainstem responses," Scandinavian Audiol., vol. 13, pp. 187-197, 1984), assume a single-stationary noise source, with the postaverage SNR increasing proportionally to the number of trials averaged. This study proposes a modified version of the Fsp statistic, the nonstationary fixed-multiple-point (NS Fmp), that can account for a discrete number of noise sources of different power, and can also be modified for weighted averaging (WNS Fmp). A new noise segmentation procedure is also proposed that dynamically partitions contiguous trials based on their noise power estimates and a series of F-tests. Results from computer simulation and real data from auditory brain stem recordings show that the NS Fmp method yields lower mean square error than do the Fsp, and that the WNS Fmp has higher receiver-operating-curve area than do the standard Fsp procedure.