Auto‐adaptive averaging: Detecting artifacts in event‐related potential data using a fully automated procedure

Abstract

The auto-adaptive averaging procedure proposed here classifies artifacts in event-related potential data by optimizing the signal-to-noise ratio. This method rank orders single trials according to the impact of each trial on the ERP average. Then, the minimum residual background noise level in the ERP data is determined at each step in the averaging process. Trials having a negative impact on the residual background noise are discarded from the averaging procedure. Simulations showed that ERP estimates obtained by the auto-adaptive averaging procedure were either better or comparable to those obtained by single trial artifact detection methods at their most optimum configuration, in particular during long duration artifacts. Experimental data from a working memory task further illustrate the effectiveness of the method.Descriptors: Artifacts, Auto-adaptive averaging, Electroencephalography, Event-related potentials Event-related potentials (ERPs) are transient electrical voltage fluctuations that the brain elicits in response to stimuli, behavioral responses, or a wide range of cognitive processes. ERPs are embedded in electroencephalographic (EEG) recordings, and are typically analyzed by averaging together segments of EEG activity (i.e., trials) that are each time-locked to the eliciting event. The underlying assumption of the averaging procedure is that the EEG signal at each trial is composed of a constant part reflecting the ERP activity and a randomly fluctuating part, constituting the concerted background activity of ongoing brain processes that are not specifically related to the event of interest. Because the randomly fluctuating background activity is by definition not correlated with the processing of the event of interest, it would theoretically average out completely should an infinite number of trials be available, leaving only the constant ERP in the remaining average. Ideally, if the ERP signal is constant, the signal-tonoise ratio of the average improves by the square root of the number of trials included in the average (beim Graben, 2001;Niedermeyer & Lopez-De Silva, 1993; see also the Appendix).ERPs are composed of a series of components, which are described by their polarity in combination with either a rank order or an estimate of the average latency of the component. For instance, P1 would indicate the first positive component and N170 would refer to a negative polarity component at 170 ms after the eliciting event. Early-latency (o200 ms after stimulus onset) components are typically related to the initial processing of sensory events and are characterized by a low amplitude (of about 1-5 mV) and a relatively high frequency. Longer latency (i.e., 200 ms and later) components are typically of a higher amplitude (10-20 mV), characterized by a low frequency, and related to higher stages of cognitive processing. In all cases, these amplitudes are significantly lower than the amplitude of the average background EEG, which is typically on the order of about 50 mV. Consequently, a large n...

Early multisensory interactions affect the competition among multiple visual objects

“…Epochs containing eye-blinks (vEOG: 150 μV/500 ms) and eye-movements (hEOG: 75 μV/20 ms; vEOG: 75 μV/20 ms) were discarded using an automated procedure (Talsma and Woldorff, 2005a). Furthermore, the signal-to-noise ration in ERP data was optimized by using an auto-adaptive procedure (Talsma, 2008). ERP data were analyzed for correct response trials only, unless otherwise stated, resulting in the exclusion of approximately 25% of trials.…”

Signal-to-noise ratio of the MEG signal after preprocessing

“…In fact, the method developed in Talsma (2008) is based on selecting singletrials according to their impact on the ERF's SNR.…”

“…The SNR has been previously employed to assess the quality of preprocessing. For instance, auto-adaptive averaging methods have been proposed to decide which epochs to reject by optimizing the SNR of the event-related potentials (ERP) (Talsma, 2008). Similarly, performing a trimmed average has been shown to be advantageous over arithmetical averages in terms of SNRs of ERPs .…”

Sequential nonnegative tucker decomposition on multi-way array of time-frequency transformed event-related potentials

“…Due to the low signal to noise ratio (SNR) in EEG data, the analysis of ERPs is usually done at the group-level [2]. The conventional data processing to obtain ERPs usually consists of five steps [1,3]: 1) filtering continuous EEG, 2) segmenting the filtered continuous EEG into single trials according to triggers of the experimental design, 3) detecting the single trials containing artifacts, 4) rejecting artifacts, 5) averaging over the single trials free of artifacts. The basic assumptions for such analysis are that data from each trial include the constant part of brain activities related to the experimental design and the randomly fluctuating part [1,3].…”