Contrasting single-trial ERPs between experimental manipulations: Improving differentiability by blind source separation

Abstract

Contrasting event-related potentials (ERPs) generated under different experimental conditions and inferring differential brain responses is widely practiced in cognitive neuroscience research. Traditionally, these contrasts and subsequent inferences have proceeded directly from ERPs measured at the scalp. For certain tasks, it is not unusual that ERPs from a subset of channels are given particular emphasis in data analysis, such as the channels displaying the maximum peak amplitude in regions of interest (“best sensors”) or channels showing the largest averaged ERP waveform differences. With the aid of a blind source separation (BSS) algorithm, second-order blind identification (SOBI), which has been recently validated for its ability to recover correlated neuronal sources, we show that single-trial ERPs from previously validated neuronal sources were more distinguishable among different experimental manipulations than the single-trial ERPs measured at the comparable “best sensors”. This suggests that by using validated SOBI-recovered neuronal sources, ERP researchers can improve the ability to detect differences in neuronal responses induced by experimental manipulations. Critically, our observations were made at the level of single trials, as opposed to the averaged ERP. Therefore, our conclusions are particularly relevant to phenomena involving trial-to-trial changes in brain activation, for example, rapid induction of brain plasticity and perceptual learning, as well as to the development of brain–computer interfaces. Similar advantages would also apply to analogous situations with magnetoencephalography (MEG).

Introduction

In cognitive neuroscience, the event-related potential (ERP) is a frequently used tool for revealing different brain activation patterns associated with experimental manipulations. One is often interested in differences between ERP waveforms and their corresponding scalp distributions induced by different experimental conditions. As high-density EEG systems have become increasingly available, it is not immediately obvious how to best analyze such high dimensional data. One approach has been to focus on a subset of sensors where ERP differences have been observed under similar experimental conditions. This approach is often used for experimental situations where ERP waveform differences are expected to be relatively focal, i.e., present at a small subset of electrode sites. Under these circumstances, statistical tests of ERP differences have often been performed on waveforms at a small number of sensors. For example, to capture the differential activations of motor cortices prior to the generation of hand movements, motor readiness potentials have been indexed using two recording electrodes, one located over the left motor cortex (C3) and the other over the right (C4) (Coles, 1989). For a given task, the specific sensors chosen have often been shown to best differentiate between experimental conditions, in this case, left and right hand movement.

Implicit in this sensor-based ERP approach is the assumption that the EEG sensors displaying the maximal signal of interest (from here on referred to as the “best sensors”) offer reasonable estimates of the underlying brain sources of interest and reasonable sensitivity to detect ERP differences between experimental conditions. One advantage of using the “best sensor” ERP approach is that the concept is intuitive and the method requires no special tools and is thus accessible to the majority of ERP researchers. Another advantage is that preselection of a few “best sensors” allows for more focused statistical tests as an alternative to the omnibus low-powered ANOVA test on data from all electrode locations. Similar to principal component analysis (PCA) or independent component analysis (ICA), which can be viewed as applying spatial filters to the EEG data, this “best sensor”-based approach can also be viewed as applying a specific kind of spatial filter. While in the case of PCA or ICA, the spatial filters are derived mathematically, in the case of the “best sensor” approach, the spatial filters are selected intuitively or heuristically. The effect of the “best sensor”-based filters is to retain signals from only the selected EEG sensors and exclude signals from all others. Because the scalp recorded EEG at each sensor contains a mixture of signals from both extra-cranial noise sources and intra-cranial neuronal sources from different brain locations, such filtering via the “best sensor” maintains the mixed nature of signals. Thus, in principle, the event-related activation from specific functional brain regions cannot be optimally indexed by scalp EEG sensor data, even when the signals are from the sensors with the largest ERP amplitude.

A variation on the “best sensor”-based analysis is taking an average surrounding and including the “best sensors”. Although such averaging offers an opportunity to remove uncorrelated sensor noise, it does nothing to unmix or isolate signals from specific brain sources of interest from those originating within other brain regions. It is critical to note that this averaging process introduces a major source of subjectivity and uncertainty concerning how many surrounding sensors and which ones to include. These decisions can be difficult as the scalp voltage map may easily reveal activation at a large number of the sensors (e.g., Figs. 3C, D). Finally, averaging across the “best sensor” and its surrounding sensors would only serve to reduce, instead of increase, the resulting signal-to-noise ratio because, by definition, the maximum is always greater than the average. Therefore, the “best sensor” approach offers better signal-to-noise ratio than the “average sensor” approach.

Blind source separation (BSS) and ICA algorithms (Hyvarinen and Oja, 2000, Vigario et al., 2000, Stone, 2002, Makeig et al., 2004, Muller et al., 2004) are potential methods for separating the mixture of EEG signals into multiple components. Some of the components recovered from these algorithms correspond to noise sources, such as artifacts (Vigario, 1997, Jung et al., 1999, Tang et al., 2002b, Joyce et al., 2004), and others to neuronal sources of interest (Vigario et al., 2000, Tang et al., 2002b, Makeig et al., 2004, Muller et al., 2004, Tang et al., 2005a). Second-order blind identification (SOBI: Belouchrani et al., 1993, Cardoso and Souloumiac, 1996, Belouchrani et al., 1997) is one BSS algorithm that can isolate ocular artifacts and sensor noise from brain signals (Tang et al., 2000a, Tang et al., 2002b, Tang et al., 2005a, Joyce et al., 2004) and separate functionally distinct but sometimes correlated brain activity, for example, activity from left and right primary somatosensory (SI) cortices following median nerve stimulation (Tang et al., 2005a). Most importantly, it has been shown that the signal-to-noise ratio (SNR) of averaged somatosensory evoked potentials (SEPs) obtained from SOBI-recovered SI components is much higher than the SNR obtained from the “best sensors” over the SI hand region (Tang et al., 2005a). This suggests that patterns of neuronal activations associated with different sensory, motor, or cognitive activations–induced by experimental manipulation–should be rendered more distinguishable when using SOBI-recovered components than when using the EEG “best sensor” signals directly.

We tested this hypothesis in a simple task domain where we attempted to differentiate single-trial ERPs induced by different stimulation conditions. We collected EEG data during an intermixed sequence of three different median nerve stimulations (left, right, and bilateral) and attempted to distinguish or classify, according to the single-trial SEPs, which of the three stimulations had been delivered during a given trial. If SOBI-recovered components isolate the neuronal responses from underlying sources better than the “best sensor” data, then one would expect that single-trial ERPs from the SOBI neuronal components should be classified with greater accuracy than those directly from the “best sensors”. To quantify the differentiability of single-trial ERPs, we used back-propagation neural networks (BPNNs) to classify the EEG data from each trial according to the stimulus condition presented. BPNNs are commonly used for pattern classification and have been previously used to classify single-trial ERPs (e.g., Pfurtscheller et al., 1996, Pfurtscheller et al., 1997). Here, we evaluated comparative performance between network classifiers trained with SOBI-recovered SI component data (SOBI component network) and those trained with the “best sensor” data (“best sensor” network). Specifically, we determined the percentage of single-trial SEPs that were correctly classified when the two different types of ERP representations were used.

It is important to point out that the goal of the present study was not to achieve optimal performance for classifying single-trial ERPs or to determine the best possible method for classification. The aim of the present study was to demonstrate the potential usefulness of SOBI preprocessing for ERP research relative to the use of the “best sensors” when contrasting the effects of experimental manipulations on brain activations at the level of single-trial. We tested a specific prediction that the SEPs of the SOBI-recovered components can serve as better indices for distinguishing among different patterns of brain activations associated with different sensory stimulations than the “best sensors” selected using information from all sensors. A variety of advanced signal processing methods, either alone or in combination, may result in further classification improvement.