A novel approach for enhancing the signal-to-noise ratio and detecting automatically event-related potentials (ERPs) in single trials

Abstract

Brief radiant laser pulses can be used to activate cutaneous Aδ and C nociceptors selectively and elicit a number of transient brain responses in the ongoing EEG (N1, N2 and P2 waves of laser-evoked brain potentials, LEPs). Despite its physiological and clinical relevance, the early-latency N1 wave of LEPs is often difficult to measure reliably, because of its small signal-to-noise ratio (SNR), thus producing unavoidable biases in the interpretation of the results. Here, we aimed to develop a method to enhance the SNR of the N1 wave and measure its peak latency and amplitude in both average and single-trial waveforms. We obtained four main findings. First, we suggest that the N1 wave can be better detected using a central-frontal montage (Cc-Fz), as compared to the recommended temporal–frontal montage (Tc-Fz). Second, we show that the N1 wave is optimally detected when the neural activities underlying the N2 wave, which interfere with the scalp expression of the N1 wave, are preliminary isolated and removed using independent component analysis (ICA). Third, we show that after these N2-related activities are removed, the SNR of the N1 wave can be further enhanced using a novel approach based on wavelet filtering. Fourth, we provide quantitative evidence that a multiple linear regression approach can be applied to these filtered waveforms to obtain an automatic, reliable and unbiased estimate of the peak latency and amplitude of the N1 wave, both in average and single-trial waveforms.

Introduction

The magnitude of event-related EEG responses is often several factors smaller than the magnitude of the background ongoing EEG. Therefore, the identification and characterization of the EEG responses elicited by sensory events (event-related potentials, ERPs) rely on signal processing methods for enhancing their SNR. The most widely used approach is the across-trial averaging in the time domain (Dawson, 1951, Dawson, 1954). The obtained waveform expresses the average scalp potential as a function of time relative to the onset of the sensory event. The basic assumption underlying this procedure is that ERPs are stationary (i.e., their latency and morphology are invariant across trials) and will therefore be unaffected by the averaging procedure, while the ongoing electrical brain activity behaves as noise unrelated to the event, and will therefore be largely cancelled out by the averaging procedure, thus enhancing the SNR of ERPs (Mouraux and Iannetti, 2008).

The cost of this across-trial averaging procedure is that all the information concerning across-trial variability of ERP latency and amplitude is lost. However, this variability may reflect important factors such as differences in stimulus parameters (duration, intensity, and location) (Iannetti et al., 2005b, Iannetti et al., 2006, Mayhew et al., 2006), and, most importantly, fluctuations in vigilance, expectation, attentional focus, or task strategy (Haig et al., 1995). Hence, the ability to obtain a reliable single-trial estimate of ERP latency and amplitude would allow exploring the single-trial dynamics between these ERP measures, behavioural variables (e.g. intensity of perception, reaction time) (Iannetti et al., 2005b) and also measurements of brain activity obtained using different neuroimaging modalities (e.g. fMRI) (Mayhew et al., 2010). Therefore, methods that explore ERP dynamics at the level of single-trials can provide new insights into the functional significance of the different processes underlying these brain responses (for a review see Mouraux and Iannetti, 2008).

Brief radiant heat pulses, generated by infrared laser stimulators, excite selectively Aδ- and C-fibre free nerve endings located in the superficial layers of the skin (Bromm and Treede, 1984). Such stimuli elicit a number of brain responses that can be detected in the human EEG both in the time domain (laser-evoked potentials, LEPs; Carmon et al., 1976) and in the time–frequency domain (Mouraux et al., 2003). LEPs are related to the activation of type-II A-mechano-heat nociceptors (Treede et al., 1998) and spinothalamic neurons located in the anterolateral quadrant of the spinal cord (Treede et al., 2003). They comprise a number of waves that are time locked to the onset of the stimulus. The largest wave is a negative–positive complex maximal at the scalp vertex (N2–P2, peaking at 200–350 ms when stimulating the hand dorsum) (Bromm and Treede, 1984). This complex is preceded by a smaller negative wave (N1) that overlaps in time and space with the larger subsequent N2 wave, and is described as having a distribution maximal over the temporal region contralateral to the stimulated side. In order to isolate the N1 wave from the N2 wave, the N1 is usually detected at the contralateral temporal electrode (Tc) referred to a frontal midline electrode (Fz or Fpz) (Fig. 1, top left panel), an EEG montage that is also recommended for recording LEPs in clinical settings (Cruccu et al., 2008, Kunde and Treede, 1993, Treede et al., 2003). Several studies have shown that the N1, N2 and P2 waves reflect a combination of cortical activities originating from primary and secondary somatosensory cortices, the insula, and the anterior cingulate cortex (Cruccu et al., 2008, Garcia-Larrea et al., 2003).

While the N2 and P2 waves are characterized by a high SNR (with a peak-to-peak amplitude of several tens of microvolts when averaging 20–30 trials) and can be easily detected in single trials (Carmon et al., 1980, Iannetti et al., 2005b), the N1 wave has a smaller SNR and is thus more difficult to detect. This difficulty is not only due to the fact that the N1 wave is generated by neural activities of smaller magnitude than those underlying the N2 and P2 waves (Cruccu et al., 2008, Treede et al., 2003). It is also due to the fact that the N1 and N2 waves (Kunde and Treede, 1993) overlap in time and space with opposite polarities, and to the fact that temporal electrodes are often contaminated by artifacts related to the activity of the temporalis muscle. For all these reasons, the vast majority of physiological (Iannetti et al., 2003) and clinical (Treede et al., 2003) LEP studies conducted in the past 30 years have relied uniquely on measures of the N2 and P2 waves to investigate the nociceptive system. In recent years, a growing number of studies have started to explore experimental modulations of the latency and amplitude of the N1 wave, and to characterize its functional significance (Ellrich et al., 2007, Iannetti et al., 2008, Lee et al., 2009, Legrain et al., 2002, Mouraux and Iannetti, 2009, Schmahl et al., 2004). Indeed, there is experimental evidence indicating that the N1 wave represents an early stage of sensory processing more directly related to the ascending nociceptive input (Lee et al., 2009), while the later N2 and P2 waves appear to reflect neural activities largely unspecific for the sensory modality of the eliciting stimulus (Mouraux and Iannetti, 2009). For all these reasons, a more systematical examination of N1 has been recommended to enhance the sensitivity of LEPs in clinical applications (Cruccu et al., 2008, Treede et al., 2003).

Hence, a technique to assess reliably the magnitude of the N1 wave has been long awaited by the scientific community, and its availability would represent a significant improvement for a more complete utilization of LEPs to explore the nociceptive system. Here we describe a novel method that combines independent component analysis (ICA) with wavelet filtering and multiple linear regression to obtain automatic and unbiased measures of latency and amplitude of ERPs, both in averaged and single-trial responses. When applied to LEPs, we show that this method provides a reliable estimate of the latency and amplitude of the N1 peak at single-trial level.