4.0 Discussion
The purpose of this study was to establish the feasibility of using the Single Trial Peaks (STP) procedure to derive amplitude and latency measures of multiple components in single ERP segments. The reliability of the STP procedure was assessed by calculating split-half and test-retest reliability coefficients for each component, and compared these to the coefficients derived from the Single Trial Windows (STW) approach. For P1, N1, P2, N2 and P3 components, the split-half coefficients obtained for STP amplitude measures ranged from .62 to .87 for Block 1 and .93 to .95 for Block 2. These reliability coefficients were generally greater than their corresponding coefficients for the STW approach that ranged from .33 to .79 for Block 1 and .59 to .95 for Block 2. The test-retest reliability coefficients for STP amplitude measures ranged from .87 to .91 and were less variable across the components than the STW approach that ranged from .60 to .94. Furthermore, the STP approach produced highly reliable latency measures at the trial level – measures that the STW approach cannot produce.
As part of the assessment of the validity of the STP procedure, descriptive comparisons between the STP and STW approaches were made for the SNR and the SEM measures. For all components, the grand means of the SNR were greater for the STP approach compared to those obtained from the STW approach. This difference in signal to noise ratio suggests that the STP procedure is better at indicating that the signals are distinct from baseline activity and are therefore valid measures of component peak amplitude. The comparisons of the grand means of the SEM showed that the two approaches were comparable and were very strongly correlated (r = .93 - .97), indicating both approaches had nearly identical degrees of measurement error across participants. The validity of the STP approach was further demonstrated by using multiple regression analyses. Four variables from each participant were used in each analysis; the participant’s mean and standard deviation of single trial amplitude measures across trials, the participant’s standard deviation of single trial latency measures, and the participant’s mean of the “noise” measures. These analyses revealed that these four variables accounted for 73% (P1) to 95% (P3) of the variance of the corresponding component derived from the individuals’ averaged ERPs.
Interestingly, when the average noise was added in the third step of the predictive models, its beta value was significant and its inclusion in the model reduced the beta weight of the standard deviation of the amplitude such that it was no longer significant in four of the five of the components examined. These results suggest a collinearity exists between the standard deviation of the peak amplitude and the amplitude variability during the baseline period, the noise measure used as the denominator in calculating SNR values. Taken together, the above results indicate that the STP approach produced highly reliable and valid peak amplitude and latency measures at the trial level.
4.1 Revisiting Assumptions Regarding Measuring Peaks in Single Trial ERPs
In EEG research, it is commonly assumed that the electrical activity in an ERP elicited by a single stimulus event is unique to the stimulus and is assumed to be invariant in amplitude and latency for each occurrence of the stimulus (Dawson, 1954; Hu et al., 2011; Quian Quiroga & Garcia, 2003). Furthermore, the elicited signal is assumed to be smaller than the electrical activity from unrelated noise sources (i.e., the elicited signal has a low SNR) and, consequently, is hidden in the mixture of activity from multiple brain sources not associated with the stimulus (Dawson, 1954; Quian Quiroga & Garcia, 2003). The STP approach was designed and implemented to be independent of these assumptions.
After filtering and segmenting the continuous EEG for a given stimulus, visual inspection of trial-level ERPs revealed multiple fluctuations in voltage, both large and small, that do not appear in the individual’s averaged ERPs. Despite the multiplicity of fluctuations, many of the positive and negative peaks in the trial-level ERP coincided to high degree with the amplitude and latency of the peaks in the averaged ERP. Thus, the general morphology of the averaged ERP was visible in almost all trial-level ERPs with the most maximal peaks (positive or negative) occurring at time periods closely related to the latencies of the successive components of the averaged ERP. As the averaged ERP is determined mathematically incorporating each time point, the general morphology of the averaged ERP must be present in each trial-level ERP to some degree. If this was not the case, the plot of an individual’s averaged ERP would begin to approximate a straight line as the number of segments in the average increases.
Using the morphology of the individual’s averaged ERP to set the time windows to select the maximal positive or negative amplitudes for successive components in each of the individual’s trial-level ERPs is the cornerstone of the STP approach. Additionally, this approach assumes and allows for the latency of the maximal peak to vary within the time window that defines the component. By considering the entire morphology rather than an isolated peak, a better understanding of how brain processing influences behavior is obtained (Gaspar et al., 2011; Hu et al., 2011; Lin et al., 2021). If change across the entire morphology is systematic rather than random, analysis of single trials can show valuable information, which could be important to understanding human function such as the state of the participant, the presence of a disability or disorder, or the ability to learn over time (Pernet, Sajda, & Rousselet, 2011; Rey, Ajmadi, & Quian Quiroga, 2015).
Thus, choosing peaks in an ERP for measurement at the trial-level based on matching an approximation of the morphology of the averaged ERP is independent of the degree of background brain activity in the ERP, though a portion of the amplitude value of the peak may represent the background activity to some degree. From the viewpoint of test and measurement framework, the best estimate of the brain “signal” elicited by the stimulus would be the mean of all trial-level peak amplitude measures and the variability of the trial-to-trial amplitude measures that represent the measurement error of the signal due to variations of background activity.
Though not a specific aim of the present study, the results of the STP approach provided evidence that the assumptions of signal invariance and low SNRs may be misleading, at least for ERPs obtained from adults while performing a speeded version of the flanker task.Related to signal invariance , the peak amplitude and latency measures of each component as defined by their maximal/minimum values in respective time windows were found to vary trial-to-trial within and across participants (Tables 2 and 3). Thus, one cannot assume that a brain’s electrical activity elicited by a unique stimulus is invariant. In addition, one cannot assume that the amplitude and latency variability represents only the “noise” from the electrical activity of unrelated brain sources as the curve fitting results suggest that as much as 20 – 25% of the variance across trials represent systematic change in brain activity that leads to performance changes in RT.
The results of the present study also question the validity of theassumption of low SNR in the EEG signal post onset of a stimulus. While the SNRs of the STP approach were lower than the traditional averaged ERP approach, the SNRs ranged from 1.6 for P1 to 2.9 for P3 and all were higher than the best SNR of the STW approach which was 1.5 for P3. These results suggest that single-trial peaks (i.e., the signal) are sufficiently large to differentiate from the noise. The findings shown in Table 4, the coefficients of variation, a measure of relative variability, were the lowest for measures from the STP approach compared to the other two approaches. The low coefficients of variation found for the peaks from the STP approach also support the conclusion that the elicited signals from stimulus are large enough to allow ERP components to be measured at the single-trial level.
Given that this study’s SNR values are at or exceed 1.6 for all components, procedures such as ICA and wavelet transform were deemed not necessary to further increase the SNR. While these procedures have been used to enhance SNR in other single-trial studies, they may also manipulate the raw signal in a manner that may increase measurement error. The data used for measurement of the ST peaks were the same segments used in producing the averaged ERP. Thus, these segments were pre-processed using the same steps and parameters used to obtain averaged ERPs, and no other manipulations of the raw signal were employed. By avoiding additional processing steps when the mean of the peak amplitudes across trials is obtained for a given component, the functionality of the component may be directly interpreted based on the literature associated with the respective component as measured from averaged ERPs. And, given that it is has less measurement error (i.e., not confounded with latency jitter) it is more likely a truer representation of the functionality of brain processing within participants than the measurements taken from an averaged ERP.