The participants were neurologically intact, healthy, adult right-hand- and right-eye-dominant males (N = 10; age range 27–39). The voluntary informed consent of the participants was obtained following US Army human use regulations approved by the Army Research Laboratory Institutional Review Board (32 CFR 219 and DoDI 3216.02).

The driving simulator consisted of a 24″ monitor, computer speakers, and a steering wheel with accelerator and brake foot pedals situated in an acoustically-attenuated room. Simulation of the visual and audio task environment was developed and rendered in real-time using SimCreator (Real Time Technologies, Ann Arbor, MI, USA). The simulation environment consisted of a long straight four-lane highway with minimal scenery (highway, roadside and horizon) and no traffic or changes in roadside stimuli, with the exception of a speed limit sign that was presented approximately every 10 min (Figure 1A). Vehicle and environmental sounds were muted, except for a tone that activated if the participants drove beyond the shoulder of the right-most lane (off the road) or across the median of the highway (into oncoming traffic lane).

The participants first acclimated to the simulation environment and vehicle dynamics during a 15-min drive where the vehicle speed was constant at 45 mph and the participants only had to control steering. Following a 5-min break, participants drove for 45 min. They were instructed to steer the vehicle to maintain the vehicle in the center of the right-most lane and adjust their speed in accordance to speed limit signs (25 mph or 45 mph). The participants were informed that a perturbation force would occasionally occur, causing the vehicle to veer left or right in a manner similar to the effects experienced when a gust of wind crosses a real vehicle (e.g., see Thiffault and Bergeron, 2003; Lin et al., 2005). They were asked to correct the vehicle by steering it back to the center as soon as possible in the event of a perturbation. The environment included perturbations that were randomly distributed in right and left directions at a rate of one per 8 s (+0–2 s) with the criteria that participants maintain the vehicle within the right-most lane boundaries for at least 8 s before the next perturbation would occur (Lin et al., 2005). The perturbation force was canceled once participants turned the steering wheel 4° in the compensatory direction. If they did not respond within a few seconds, the force would angle the vehicle into the left or right side boundaries and trigger the warning tone. Every 15 min, the experimenter talked to the participant over an intercom and asked the participant to verbally report their sleepiness rating from 1 (extremely alert) to 9 (extremely sleepy, fighting sleep) when prompted. In the present article, our analysis focuses only on the data during the 45 min drive.

Although the participants drove continuously, three metrics were computed around each perturbation event. The LD was measured in meters and computed as the absolute value of the vehicle’s position from the center of the participant’s traveling lane (Figure 1B, top). The HE was measured in degrees and computed by determining the angular difference between the vehicle’s current course and an infinitely straight line that is parallel to the road edge (Figure 1C, top). Finally, the integrated steering response (ISR) was the integral of the steering wheel position from the time of response to 3 s after the perturbation started (Figure 1D, top). Each calculation occurred when the participant’s steering wheel response was ≥4° in the direction counteractive to the perturbation force. The time elapsed between the initiation of the perturbation and the response determined the RT (circle, Figures 1B–D, bottom). Simple t-tests showed no differences (all p > 0.1) in the steering responses, HE, RT or LD as a function of leftward or rightward perturbations. Therefore, in our subsequent analyses, we collapsed across right and left perturbations, and this decision did not impact the brain data analysis which examined the 3 s before perturbation onset. Finally, to compare measures across subjects, each driving performance measure was z-scored separately within each subject so that across subject variability was normalized.

EEG data were acquired using a BioSemi system (Amsterdam, Netherlands). Continuous recordings were sampled at 2048 Hz and acquired from 64 standard scalp locations (10–10 system; Chatrian et al., 1988) referenced online to the Common Mode Sense (CMS) electrode. Six additional channels were also recorded with online reference to CMS: left/right mastoids (M1/M2), two horizontal (left/right; HEOL/HEOR) and two vertical (upper/lower right eye; VEOU/VEOL) electro-ocular electrodes. EEG data were synchronized with the driving simulator CPU’s via an Arduino-based system.

All signal processing was applied using EEGLAB (ver 11.0.0.0b; Delorme and Makeig, 2004¹) and in-house code using Matlab (R2012a) on a 64-bit Linux operating system. Following established preprocessing procedures (Onton et al., 2006), all electrodes (64 EEG + 4 EOG) were first re-referenced to the average of the two mastoid electrodes (M1 and M2) before being bandpass filtered between 1 Hz and 50 Hz using a zero-phase finite impulse response filter (EEGLAB function pop_eegfilt). The EEG data for each participant was epoched into non-event-locked, adjacent (non-overlapping), 500-ms windows, and automatic artifact-rejection algorithms were applied to the epoched data using EEGLAB function pop_rejmenu with the following rejection criteria: (1) amplitude threshold (> ±100 μV); (2) joint probability (>5 SD); (3) abnormal trends (max slope 75 μV/500-ms epoch; R-squared limit 0.3); and (4) kurtosis (>5 SD). Finally, we visually inspected all epochs and confirmed automatically-tagged epochs containing artifacts or manually accepted or rejected epochs if incongruent with visual inspection.

We applied infomax independent component analysis (ICA; Bell and Sejnowski, 1995; Lee et al., 1999) and obtained spatially fixed, maximally temporally IC processes (Makeig et al., 1996) for each participant. A single model decomposition was used with default extended-mode training parameters (EEGLAB function runica). We then obtained dipoles for all ICs using EEGLAB plugin dipfit and selected features for clustering ICs using EEGLAB function pop_preclust (spectra = 1, ERSP = 1 and dipole models = 10). By default this technique uses the Four-Shell spherical head model. Clustering was performed using the k-means algorithm (EEGLAB function pop_clust) with number of clusters set to 19 and outliers set to >2.5 standard deviation from cluster centers.

Based on our a priori hypothesis, our analysis focused on two of the identified clusters in left and right parietal regions. The algorithm assigned 10 total ICs to each of the parietal clusters (Figure 2), including one component per participant whose dipole was proximal to parietal cortex so no manual adjustments were made to the output clusters. For each IC in each parietal cluster, we generated epochs around each perturbation event (−3000 ms to 3000 ms) and z-scored the power on a trial-by-trial basis using a wavelet decomposition (Morlet) to obtain average dynamic changes in spectral amplitude within the alpha band (8–13 Hz). This procedure aligns with previous research that demonstrated less sensitivity to noisy trials by using the entirety of the trial epoch period, instead of the traditional pre-stimulus period, when computing spectral power estimates (Grandchamp and Delorme, 2011). This normalization procedure (using the newtimef function in EEGLAB) also ensured common units of standard deviation so that we could directly compare behavior to neural data. The regression analysis, then, focused on the pre-perturbation period and used the averaged z-scored values of alpha power from −3000 ms to 0 ms. As confirmatory analyses, we also verified that the inclusion of the post-perturbation period in the normalization procedure did not bias the results by comparing the correlation between pre- and post-perturbation values (R² = 0.1 for left hemisphere and R² = 0.1 for right), and we did not find significant correlations between the alpha activity during the post-perturbation and any of the four metrics of driving performance (p > 0/05).

We used a GLM to investigate the relationship between driving performance and alpha activity in right (R_α) and left (L_α) parietal dipole IC sources. Four separate models were run for the four driving metrics of interest (Behavior): LD, RT, HE and steering response. We controlled for time-on-task effects that may also modulate alpha activity and changes in vehicle-road kinematics (Otmani et al., 2005; Wascher et al., 2016) by including a parameter for time elapsed from the beginning of the experiment to the time of the trial measured in seconds (Time). In the equation, the * denotes that all main effects, pairwise effects, and three-way interactions were modeled simultaneously.

For each GLM, a backward selection procedure was used to select the most parsimonious model that was able to account for the variability in behavioral performance across 1896 trials. First, the model was fit with coefficients for the three main effects (Right, Left and Time), the pairwise effects (Right-Left, Right-Time, Left-Time), the three-way interaction (Right-Left-Time) and the intercept. The coefficients were then sorted by p-values, and the coefficient with the largest, insignificant p-value (p > 0.05) was eliminated from the model. If that coefficient represented a main effect, all higher order interaction terms containing that effect were eliminated. This procedure continued until all remaining coefficients were significant (p < 0.05) and identified terms with a significant relationship to the driving behavior, or the procedure terminated when no terms remained in the model which indicated no relationship with the driving behavior. To control for multiple comparisons, we performed false discovery rate correction (Benjamini and Hochberg, 1995). As a confirmatory analysis, we also examined within-subject variability by including a subject-specific term in the regression model, but the direction and significance of the identified coefficients in Table 1 remained the same with only small fluctuations in magnitude.

The average number of total perturbations across participants: 189, standard deviation (SD): 29. As visualized in Figure 1, we computed four performance metrics around each perturbation event, LD, HE, RT and steering response. These metrics were calculated at the onset of the participant’s steering response to counteract the perturbation force. The average RT across all participants: 1.16 s, SD: 0.46 s; the mean LD: 0.373 m, SD: 0.33 m. The average HE at the time of response: 23.49°, SD: 29.28°; and the average ISR: 6.71°, SD: 4.88°.

We computed a general linear model to examine driving performance as measured by LD as a linear combination of left and right parietal alpha power, time-on-task effects, and the interaction of these variables. In the initial model fit (F_(7,1888) = 2.6, corrected p-value = 0.02), prior to the backward selection procedure, the p-value for all terms were larger than 0.1 except for the main effect for right parietal alpha activity and the three way interaction term (see Table 1). After the model selection procedure; however, the only term that remained in the model was the right parietal alpha activity (coefficient = 0.047 with a 95% confidence interval of [0.002–0.09], p-value = 0.038). The relationship between right alpha power and LD for all subjects is plotted in Figure 3.

Following the same procedure used for LD, we computed a general linear model to examine driving performance as measured by the vehicle HE as a linear combination of left and right parietal alpha power, time-on-task effects and the interaction of these variables. In the initial model fit (F_(7,1888) = 6.76, corrected p-value < 0.01), prior to the backward selection procedure, the p-value for all terms were larger than 0.1 except for the main effect for left parietal alpha activity and the main effect for time-on-task (see Table 1). After the model selection procedure, however, the same two terms remained in the model, left parietal alpha activity (coefficient = 0.09 with a 95% confidence interval of [0.046–0.13], p-value < 0.01) and the time-on-task (coefficient = 0.11 with a 95% confidence interval of [0.07–0.16], p-value < 0.01). The relationship between left alpha power and HE for all subjects is plotted in Figure 4A, and the relationship between time-on-task and HE for all subjects is plotted in Figure 4B.

The last model we examined was the relationship between the ISR, parietal alpha activity, and time. In the initial model fit (F_(7,1888) = 6.5, corrected p-value < 0.01), prior to the backward selection procedure, the p-value for all terms were larger than 0.1 except for the main effect for left parietal alpha activity and the main effect for time-on-task (see Table 1). The same two terms remained in the model after the model selection procedure, left parietal alpha activity (coefficient = 0.07 with a 95% confidence interval of [0.052–0.114], p-value < 0.01) and the time-on-task (coefficient = 0.11 with a 95% confidence interval of [0.065–0.154], p-value < 0.01). The relationship between left alpha power and the ISR for all subjects is plotted in Figure 5A, and the relationship between time-on-task and ISR for all subjects is plotted in Figure 5B.

The last model we examined was the relationship between the RT, parietal alpha activity, and time. In the initial model fit (F_(7,1888) = 5.13, corrected p-value < 0.01), prior to the backward selection procedure, the p-value for all terms were larger than 0.07 except for the main effect for left parietal alpha activity and the main effect for time-on-task (see Table 1). The same two terms remained in the model after the model selection procedure, left parietal alpha activity (coefficient = 0.1 with a 95% confidence interval of [0.059–0.148], p-value < 0.01) and the time-on-task (coefficient = 0.07 with a 95% confidence interval of [0.023–0.113], p-value < 0.01). The relationship between left alpha power and the RT for all subjects is plotted in Figure 6A, and the relationship between time-on-task and RT for all subjects is plotted in Figure 6B.

The right parietal component cluster demonstrated a significant, positive relationship with LD. Previous work has demonstrated that activity in right parietal cortex supports visuospatial processing and orienting (Whitehead, 1991; Rushworth et al., 2001; Foxe et al., 2003; Thut, 2006), and additional research has refined its role, suggesting that the right hemisphere is more closely aligned to events with higher temporal predictability, increased arousal, or heightened vigilance (Fernández and Siéroff, 2014). In line with these findings, right parietal alpha activity increased when the participant was slow to correct for the recent wind perturbation and allowed the amount of LD to increase. Increased pre-stimulus alpha power has been linked to decrements in subsequent performance (Hanslmayr et al., 2011), and in the context of driving, it has been interpreted as decreased arousal with impaired vigilance (Lin et al., 2012). Our findings suggest that the right parietal cortex reflects visuospatial processing that is critical for tracking vehicle kinematics for the central task (maintaining a zero LD), and fluctuations in alpha may capture fluctuations in attention that are critical for driving performance. Interestingly, recent clinical work supports this interpretation by demonstrating the patients with right hemisphere lesions have significantly more lane crossing incidents that those with left hemispheric lesions, and furthermore, this study found that left hemisphere patients demonstrated longer RT for braking responses than those with right hemisphere lesions (Park, 2015). These clinical findings nicely complement our results from healthy participants.

The left parietal component cluster demonstrated a significant, positive relationship with HE and steering response. Previous behavioral research has shown that HE is closely related to upcoming steering responses (Hildreth et al., 2000; Field et al., 2007). Our results align well with previous research that suggested that left parietal regions may direct attention for upcoming movements (Rushworth et al., 2001) as well as studies where preparatory attention is modulated by the left hemisphere (Fernández and Siéroff, 2014). Collectively, our current results suggest that left parietal alpha activity relates to processes which monitor the current heading of the vehicle to determine the timing and magnitude of the response for subsequent corrective action.

These results also align with previous research that uncovered differential network activity related to adaptive and stable task control (Dosenbach et al., 2007). Their study uncovered a frontoparietal network that demonstrated error-related activity on a trial by trial basis, and a secondary network that had sustained activity throughout the task. Our results are similar. Activity in the left parietal cluster related to HE, RT and coordinated steering control—all of which were dynamic over the course of the experiment. Conversely, the right parietal region had sustained activity related to the critical task parameter, LD, which did not change over the course of the experiment. These findings suggest that parietal activity may reflect a dynamic control strategy to maintain critical performance when executing a visual-motor task.

Across the models for the four driving performance metrics, LD did not show a time on task effect, while both HE, RT and steering response increased over the course of the long drive. Previous research has suggested that humans are exceedingly vigilant against infractions of the central task requirement (Hockey, 1997; Grier et al., 2003; van der Linden et al., 2003), and in our study, the participants were instructed to keep their vehicle in the center of the traveling lane. In support of this previous literature, our results revealed that LD did not show a main effect of time, indicating that participants successfully minimized performance decrements in their primary task.

In contrast, the models for both HE and steering response revealed a main effect of time. The behavioral coupling between these two driving metrics has been previously shown suggesting that they may be incorporated in a complex feedback control system wherein the HE is perceived and transformed into a steering response. In support of this, behavioral research has found that the HE is more closely linked to steering responses than to LD (Hildreth et al., 2000; Wallis et al., 2002; Cloete and Wallis, 2009; Li and Cheng, 2011). In our study, we extended these results and observed that both driving metrics increase over the duration of the experiment. When construed as a change in the control parameter, the increase in HE over time suggests that the brain permits increased error in this regulatory dimension, but in order to maintain the LD near zero for the central task requirement, the participant utilized larger steering wheel corrections. This paired increase in HE and steering response is analogous to uncontrolled manifold hypotheses (Scholz and Schöner, 1999; Todorov, 2004) wherein a number of different heading/steering corrections over time can lead to a small, near zero LD. Combined, our results suggest that the brain utilizes HE to estimate steering responses, and as time on task effects negatively impact driving performance (Matthews and Desmond, 2002), the increased tolerance for larger HEs leads to a concomitant change to larger steering wheel corrections. This is supported by the observation that RTs increased over the course of the experiment which may suggest that monitoring processes degraded over the course of the experiment and resulted in larger heading responses and subsequent increases in steering responses. Our data cannot be directly interpreted as an explicit feedback control model since we do not model the dynamics of this relationship (e.g., steering responses over time), but this pattern can be interpreted in terms of a feedback control system that allows varying levels of error in control parameters while maintaining accuracy in the central visual-motor task. More specifically, right parietal activity in our study was associated with monitoring of LD (the central, critical task requirement), while left parietal activity associated with feedback control perhaps in the computation of the control signal. Future research can target experimental designs to test these interpretations.