Pupil-linked arousal correlates with neural activity prior to sensorimotor decisions

Objective. Sensorimotor decisions require the brain to process external information and combine it with relevant knowledge prior to actions. In this study, we explore the neural predictors of motor actions in a novel, realistic driving task designed to study decisions while driving. Approach. Through a spatiospectral assessment of functional connectivity during the premotor period, we identified the organization of visual cortex regions of interest into a distinct scene processing network. Additionally, we identified a motor action selection network characterized by coherence between the anterior cingulate cortex (ACC) and dorsolateral prefrontal cortex (DLPFC). Main results. We show that steering behavior can be predicted from oscillatory power in the visual cortex, DLPFC, and ACC. Power during the premotor periods (specific to the theta and beta bands) correlates with pupil-linked arousal and saccade duration. Significance. We interpret our findings in the context of network-level correlations with saccade-related behavior and show that the DLPFC is a key node in arousal circuitry and in sensorimotor decisions.


Introduction
Navigating our environment requires us to make continuous sensorimotor decisions in complex scenarios.For example, driving a car in a new city requires us to interpret the external world so that we can understand the traffic rules, where to turn, and plan several motor actions in advance, accounting for past actions and their consequences.
The arousal system, specifically pupil-linked states related to locus coeruleus (LC) activity [1,2], has been shown to play an important role in regulating the influence of incoming information on beliefs about a dynamic world [3].The mechanism of arousal modulation is still a relatively unexplored area [4].However, researchers have previously shown the interaction between the vagus nerve and LC activity to be critical to the treatment of clinical conditions [5].This relationship between arousal and performance has been shown in a variety of conditions [6,7] through the Yerkes-Dodson Law [8].It is believed that by comparing the expected to the actual sensory events that are experienced, the central nervous system can monitor task progression, detect performance errors, and quickly launch appropriate task-protective corrective actions as needed [9,10].
Studies have shown a link between arousal and performance in stressful scenarios like driving [11,12] and boundary-avoidance tasks (BATs) [13].In BATs, pilots must avoid collisions with the ground or avoid stall angles through boundary tracking, which triggers the fight or flight response and affects the integration of sensory evidence [14].While BATs were initially studied in pilots, road driving is a similar scenario that also requires multiple cognitive functions, including perception, attention, motor control, and decision-making [15].Sensorimotor decisions in the context of driving are generally characterized by three actions: acceleration, braking, and steering.While acceleration and braking have been studied in discriminative studies [16], steering has been shown to be a reliable, continuous corollary of proactive brain states relevant for navigation [17][18][19][20].Researchers have identified three stages characterizing action selection in realistic, driving-related scenarios.A high level, strategic, exploratory period where we process information about the environment to select a route and traffic rules, a middle motor preparation period where the brain maps the sensory information into an ideal course of action, and the lowest level characterized by action execution and perceptual processing [21][22][23][24][25].
Methods such as eye tracking (ET), electroencephalography (EEG), and electrocardiography (ECG) have been used to demonstrate that decision certainty influences several ocular and autonomic measures, including phasic pupil diameter [26] and high-frequency heart rate variability (HF-HRV) [27], which in turn are relevant to motor action selection.However, to our knowledge, there have been few studies on the activity during exploratory and action selection periods of the kind that are prevalent in complex motor movements, such as driving, with simultaneously collected neural and autonomic measures.
Current theories on action selection emphasize the role of the dorsolateral prefrontal cortex (DLPFC), the premotor cortex, and the anterior cingulate cortex (ACC) for planning motor executions [9,28,29].In humans, functional connectivity between most motor-related areas has been shown to increase during motor action selection periods [30].Lesion-induced disruption of sensorimotor networks has been shown to impact decisions in stroke patients, suggesting that network-level analysis of premotor activity may yield better prediction of behavior and, ultimately, clinical outcomes [31].In this study, we introduce the term 'motor action selection network' to denote the DLPFC and ACC functional network, which appears to play a role in higher-level scene processing relevant for motor action selection.Brain oscillatory power in the beta-band (13-30 Hz) has been shown to decrease during the preparation, and the execution of voluntary movements [32].Neural activity in the beta band of the cingulate cortex has been shown to be sensitive to errors, particularly while driving [17,33].EEG data have also shown that there is a significant increase in theta and delta activity [34], however the literature in this area has been inconsistent.
Recent work has proposed several noninvasive methods to monitor arousal state and neural drivers of sensorimotor decisions.The eyes play a critical role in scene processing and driving motor movements [35][36][37], with measures such as eye closure, blinking and saccades providing real-time indices for arousal while driving [38,39].However, the reliability of such methods varies greatly with brightness in the environment, necessitating new methods to control for luminance effects on the eyes [40] and measures of vagally-mediated arousal.The striate (V1) and V2 regions of the visual cortex have displayed distinct functional properties.Their anterior and posterior subregions correspond to representations of the peripheral and central visual fields, respectively [41].Moreover, these regions have been implicated in a broader scene processing network responsible for the low-level processing of visual features [42].In our context, we introduce the term 'scene processing network' to denote the V1 and V2 functional network, which seems to play a role in lower-level feature processing.
In this study, we investigated whether cortical networks predictive of motor actions are correlated with arousal and eye-linked behavioral measures (figure 1).We expected regions of interest (ROIs) related to low-level visual processing, exploratory activity, and those involved in arousal pathways to be relevant in the prediction of motor intensity, specifically absolute steering deviation.These areas were also expected to correlate with pupil-linked arousal, HF-HRV, and saccade-related behavior.Participants actively drove through a virtual city environment in a BAT.We identified steering events using thresholded, absolute steering deviation (figure 2(B)) and epoched 1000 milliseconds of data leading to each event (premotor period, figure 2(C).After source reconstruction, we selected 13 ROIs relevant for action selection and execution, and we modeled absolute steering deviation as a function of oscillatory EEG activity.In a hierarchical design, we used significant ROIs to describe functional connectivity between ROIs, showing the presence of a lowerlevel perceptual processing network characterized by visual cortex activity and a higher-level exploratory network associated with DLPFC and ACC activity.We show that activity in the exploratory network correlates with pupil-linked arousal, calculated using luminance-adjusted pupil diameter and HF-HRV, a measure of parasympathetic vagal activity.We also report a weak correlation with saccade-related eye behavior.Our results show that neural oscillatory power associated with motor actions in complex environments is partly influenced by the arousal pathway.

Experimental paradigm
All experiments were approved by Columbia University's Institutional Review Board, and all participants gave written informed consent to participate in this study.Ten subjects completed a BAT in a simulated city environment.The task required them to steer a car left or right to avoid boundaries and reach a designated target location.The task was presented using a virtual reality (VR) headset (HTC Figure 1.Overview of the present study (A) EEG channel data was transformed into cortical current source density (CSD) via eLORETA.CSD was averaged in 13 areas of interest, and 5 regions of interests (ROIs) were selected for subsequent analyses: ventral/dorsal anterior cingulate cortices (vACC, dACC), visual areas 1 and 2 (V1, V2), and the dorsolateral prefrontal cortex (DLPFC).(B) We studied the functional connectivity of significant ROIs (at p < 0.05 level, * ) whose oscillatory power was predictive of motor action.We expected the organization of these source regions and correlations with pupil-linked arousal, heart rate variability (HRV), and saccade-related behavior.VIVE Pro Eye) to create a realistic and immersive environment.
We varied the task difficulty on a trial-by-trial basis by adjusting a visual noise opacity parameter.This parameter was meant to simulate the type of white, 1/f noise found in visual search tasks and was perceived by participants through the density of fog prevalent in the city environment.
In the experiment, participants operated a virtual vehicle through a city simulation with boundaries on the road's left and right sides.They were instructed to use a physical steering wheel, accelerator, and brake pedals to avoid colliding with the boundaries.To incentivize participants to complete the task quickly, they were informed that they could earn a cash bonus based on the number of completed trials, with the bonus amount resetting to $1.00 every block.They were informed that hitting the boundaries would result in damage to the vehicle.The task was designed such that five crashes at maximum speed (50 mph) during a block would deplete the bonus amount for that block.The bonus amount was reduced by a maximum of $0.20 per crash event, depending on the acceleration of the car upon impact with the road boundary.

Motor actions
We identified motor actions (figure 3(A)) through a simple post hoc peak and trough-detection technique on the steering wheel channel.Using a nonoverlapping, look-behind window of 750 ms, we assured that the peak we encountered was the true peak in steering wheel activity.The intensity of motor action is defined by the absolute value of the steering wheel deviation (in degrees, figure 2).
In the subsequent explanations of our analyses, we will refer to the premotor period, the 1-second window that precedes a motor action as a 'trial.'A total of 5126 trials were identified with this method.The start of each trial was the beginning of the premotor period, which was identified based on the detected motor action.The absolute steering deviation characterizes the intensity of the motor action for each trial.

Isolating arousal effects on pupil diameter
Before studying the effect of pupil-linked arousal on motor actions, we sought to remove the luminancedriven effects on pupil diameter (figure 4) using responses from a controlled exposure period.We found a significant effect of opacity on pupil diameter for all participants (supplementary table 1) using responses from controlled exposure periods.
We used the participant-level coefficients for residualizing the arousal-driven changes in pupil diameter after computing the expected change in pupil diameter from the exposure period and subtracting this value from the observed change in pupil diameter after opacity changes.We expect that subsequent analyses using pupil diameter thus capture mostly pupillinked arousal effects.

DLPFC, ACC and visual cortex activity is predictive of motor action
We found a significant effect of oscillatory power in several regions in the premotor period on the motor intensity (table 1) in our linear mixed effect (LME) analysis.In the areas where we expected higher-level scene processing, such as the ACC and DLPFC, the Overview of how motor events were detected and premotor period is defined.(A) We simultaneously collected neural data from EEG, autonomic measures using ECG, eye movements and pupil dynamics using a VR-headset embedded eye tracking system, and motor actions using a steering wheel.(B) Participants (n = 10) performed three virtual reality driving task sessions, requiring boundary avoidance under time pressure and changing visual uncertainty.Their motor actions were recorded from the steering deviation as they were navigating a city environment.We analyzed direction-independent (i.e.absolute) steering deviation).Motor actions belong to a global trial with a set level of visual fog (opacity) in the environment that participants drove in.(C) The start of each motor action was marked using a peak detection method on the steering wheel data since this was most relevant to navigating the BAT.The premotor periods of interest were a fixed, 1-second interval before each event, and the intensity of the motor activity was determined by the post-event steer angle.Blue and red circles indicate events with low and high motor intensity, respectively.
predictive activity was lateralized to the left hemisphere.DLPFC activity in the alpha band was the strongest predictor (β = −0.245,p < 0.001), while significant ROIs in the visual cortex (V1 and V2) were the weakest predictors (β = 0.080 and −0.072 respectively, p < 0.05).We did not find significant contributions of other areas in predicting motor action.
The results of the sliding estimator decoder are shown in figure 1(A).Notably, due to the high degree of colinearity between activity in the ROIs, the performance of most regions assessed is similar (R 2 > 0.116).

ROIs predictive of absolute steering deviation form distinct networks
We studied statistical interdependencies between significant ROIs using a phased-lag index (PLI) connectivity approach.The results displayed in figure 5 show task-dependent connectivity between V1 and V2 for alpha, beta, and gamma bands in the premotor window, as well as between DLPFC and vACC for the beta band.The alpha band also showed connectivity between vACC and the visual cortex and between dACC and vACC in the beta band.Furthermore, the gamma band produced low connectivity between non-visual cortex regions, which is the only case with low PLI values between the frontal and occipital regions.
We assessed the statistical significance of difference scores in PLIs between observed data and after random phase adjustment [44].We performed false-discovery rate-based correction, using the Benjamini-Hochberg method [45], for multiple comparisons.Connectivity patterns across all bands were found statistically significant: theta (Mann-Whitney U = 86, p < 0.01), alpha (Mann-Whitney U = 48, p < 0.01), beta (Mann-Whitney U = 96,  (A).Whereas the actions in (A) were used for event locking, the intensity of motor actions attributed to each event was binned using the absolute value of the steering deviation (i.e.'final' steering wheel degree), with respect to the start of the event, after 750 msecs.(E) Distributions, by-participant, of the primary outcome of interest, the absolute steering deviation, characterizing the magnitude of motor decisions.p < 0.01), gamma (Mann-Whitney U = 208.5,p < 0.01).
This analysis confirmed connectivity between the left and right visual cortex and vACC in the alpha band.Planned motor action may partly be influenced by the connectivity between the V1-V2 scene processing network, which has shown to be a large neuronal network with a high degree of synchronicity [46], but connectivity in the higher frequency bands validate the presence of a distinct vACC, and DLPFC attention shifting network [47].

Connectivity patterns between collision events, but not visual opacity, were significantly different
Because pupil-linked arousal in our task can be influenced by the visibility of the road and collision events, we conditioned trials in two ways: by the participantlevel mean visual opacity value and by detected collision events.Mean-splitting resulted in 2551 and 2516 'low' and 'high' density trials for computing differences in connectivity (figure 5(B) top).We found no significant differences between PLIs across all bands (U > 324, p > 0.11).
The number of collisions ranged from 16-258 (mean = 98.4,SD = 78.7)across participants, with a total number of 983 trials with collisions.Using an equal number of randomly-selected trials from non-collision trials, we computed differences in PLIs (figure 5(B) bottom) by band.We found a significant difference in connectivity patterns in the alpha band (Mann-Whitney U = 271.5,p = 0.02).

DLPFC and ACC activity is correlated with pupil-linked arousal and eye behavior
We sought to explore whether, as expected, saccaderelated behavioral measures and arousal levels correlate with DLPFC and ACC activity (supplementary figure 2).Saccade behavior has previously been shown to strongly correlate with DLPFC activity [48,49].We found a small, significantly negative correlation with DLPFC in the beta band (ρ = −0.07,p < 0.01) and a positive correlation between saccade duration and dACC activity in the theta band (ρ = 0.08, p < 0.01).
We report a high degree of correlation between DLPFC and ACC and arousal-linked measures, specifically luminance-adjusted pupil diameter and root Table 1.Linear mixed effects results: predicting motor intensity from oscillatory power (θ: 4-8 Hz, α: 8-15 Hz, β: 15-32 Hz, γ: 32-55 Hz) from source regions of interests (ROI)s in visual cortex areas (V1, V2), ventral (vACC) and dorsal (dACC) anterior cingulate cortices, and dorsolateral prefrontal cortex (DLPFC).LH and RH refer to the left and right hemispheres, respectively.A total of 5126 trials were included across participants, and each participant's intercept was modeled as a random effect.MNE coordinates refer to the source space coordinates from the MNE-Python software [43]  mean square of successive differences (RMSSD).

Neural networks involved in active, sensorimotor decision-making
Our LME results reaffirm the role of several brain regions in motor action selection and decisions.Notably, we report a distinct, lower level, 'scene processing' network in the occipital regions (V1, V2), which was a weak, significant predictor of absolute steering deviation.Erla et al [50] reported that execution of a visual and combined visuotactile task, but not a tactile-only task, induced activation in the  occipital region, suggesting the visual cortex may be involved in low-level visual processing relevant for motor execution.In fact, V1 activity in the gamma band was the only significant, yet weak, correlate of trial opacity (ρ = −0.05,p < 0.05, table S2).
Our findings affirm past findings on the role of ACC and DLPFC in motor action and planning.The ACC has been shown to be involved in evaluating action outcomes and adapting behaviors [51] in addition to cognitive control during motor tasks in rats [33].Specifically, Andreou et al [52] reported an increase of high-beta oscillations in response to feedback from positive reward in the context of a gambling task.In our study, we find increased beta vACC activity prior to high-intensity motor actions which have a larger effect on reward.The DLPFC has also been shown to contribute to efficient motor planning and execution through coordination of actions that impact reward [53].
We found that not only were some of our hypothesized ROIs significant predictors of action selection, but these regions could also be grouped into two distinct networks.Specifically, we found a prominent V1-V2 network across all frequency bands above 15 Hz (as shown in figure 5), which has been previously observed in macaques despite significant changes in stimulus and time-dependent gamma frequency [54].In a study of face processing in humans, Ioannides et al [55] reported similarly robust V1-V2 connectivity patterns in healthy adults, using mutual information analysis, which showed that MEG activations were not localized to a single area but were lateralized to the right hemisphere across multiple processing stages.Our results are consistent with these findings.
Additionally, we highlight the interconnectivity among vACC, dACC, and DLPFC in the beta band, forming a high-level 'motor action selection' network.This connection aligns with the findings of Kondo et al [47], underscoring the necessity of ACC-DLPFC communication for attention shifting-a notion underscored by our connectivity outcomes.The error-sensitive ACC's interaction with the DLPFC, recognized for its role in implementing performance adjustments within goal-oriented scenarios [56], may be driving the observed connectivity, particularly in the context of our BAT.While PLI between vACC and dACC is relatively weaker, the presence of connectivity between the regions is not surprising in the context of existing literature showing high-beta oscillations are affected by aspects of reward-related stimuli [52].Our findings support the notion of a frequency-specific, context-dependent modulation of the reward system, specifically in the context of BATs.
We observed a significant increase in connectivity involving the DLPFC, dACC, and V2, with boundary collisions (figure 5(B) bottom).We also observed a significant decrease in connectivity within the scene processing and motor action selection networks in the alpha band, which was lateralized to the right hemisphere.Interestingly, we did not find this difference in connectivity when examining the influence of visual opacity, which we expected to affect connectivity related to pupil-linked arousal.Previous research has demonstrated the co-activation of the DLPFC and ACC in error-related cognitive control [57,58].Our study provides evidence of the visual cortex's involvement in cross-network communication during premotor periods with increased sensorimotor errors.Our results support the hypothesis that the ACC detects errors that the DLPFC uses to associate errors with relevant task stimuli [57].In our case, the visual cortex may be processing relevant stimuli in the scene for sensorimotor decision-making, as evidenced by cross-network connectivity.

Arousal correlates of neural activity predictive of motor actions
Pupil diameter has been extensively linked with engagement [59], and our results support this finding in the context of driving in VR (ρ = 0.09, p < 0.01, table S2).In our study, a decrease in pupil diameter was accompanied by increased lateralized DLPFC activity (supplementary figure 2).In a study of the causal role of DLPFC on pupil diameter, Allaert et al [60] report pupil dilation following an active, anodal transcranial direct current stimulation (tDCS) stimulation of DLPFC when confronted with negative emotional images, and a decreased pupil diameter following tDCS stimulation of right DLPFC.While our corollary result is in the opposite direction, our finding agrees with the general idea of lateralized, task-dependent effects of DLPFC activity on pupil diameter.
While pupil-linked arousal is a marker of LC activity and can inform us how humans respond to changes in the environment, RMSSD is a commonly used index of parasympathetic nervous system reactivity and a marker of vagal activity.In fact, vagal activity has recently been shown to precede neural dynamics and correlate to the reported level of arousal [61].Lesions in the LC have been shown to limit the effect of vagal nerve stimulation for the treatment of epilepsy in rats in a highly connected pathway [5].For this reason, we study the effect of vagally-mediated arousal and pupil-linked arousal separately.
One of our key findings is that pupil-linked arousal negatively correlated with parasympathetic activity as measured through RMSSD (ρ = −0.09,p < 0.01).Although our observed effect is small, our findings are consistent with past work on driving in high-stress scenarios [62].The inverse relationship between RMSSD and luminance-adjusted pupil diameter can be explained through the established, underlying mechanism which controls arousallinked pupil dilation.Pupil dilation has been shown to be controlled by parasympathetic and sympathetic components, including inhibition of the parasympathetic-controlled pupillary sphincter by the Edinger-Westphal nucleus and through direct sympathetic activation of the dilator muscles [63].
We examined the relationship between neural activity in brain regions relevant to sensorimotor decision-making and autonomic state, as measured by pupil diameter and RMSSD.Previous research has established a link between ACC activity and pupil diameter [2], but our study extends this by investigating regions predictive of motor action.Our results showed that the DLPFC had a negative correlation with both pupil diameter and RMSSD, while the dACC had a positive correlation with both measures (supplementary figure 2).These findings are consistent with prior literature indicating that the DLPFC tracks arousal [64] and that the dACC supports the generation of autonomic states associated with cardiovascular arousal during motor behavior [65].

Physiological markers of motor decision-making
We hypothesized that in addition to arousal state, neural activity driving motor decisions would be jointly associated with saccade-related activity.While we found only a slight correlation in the 1-second premotor interval between the DLPFC, dACC, vACC, and saccade duration, our findings reinforce neuronal and clinical studies highlighting the role of the DLPFC and ACC on saccade behavior [66,67].

Limitations
Our results indicate that the oscillatory power of certain hypothesized ROIs can predict motor action intensity.However, we have identified limitations that should be considered in future studies.Firstly, our prediction of motor action intensity relies solely on EEG data.While our study did not primarily focus on accuracy, it is important to note that using unimodal data for prediction has a general limitation in its capacity to fully capture the planning necessary for executing our target motor actions.To overcome this limitation, integrating data from other modalities that have shown correlations with the ROIs and were used to characterize the premotor period could improve the accuracy of our target representation.
Another potential limitation in our study concerns the trial selection process for motor action prediction.We identified trials indirectly through a post hoc peak detection technique.The size of the lookback window used to determine motor actions and the duration of the premotor period could potentially impact our overall prediction accuracy.To mitigate this limitation, we can explore different lookback window sizes and premotor period durations.
Lastly, it is important to note that our experiments were conducted within a controlled virtual scenario.In real-world settings, numerous factors and distractions can potentially influence how individuals perceive their environment and make decisions.Therefore, results obtained from the VR scenario may not be an ideal representation of real-world scenarios.Consequently, conducting similar tests with real-world data, rather than relying solely on VR, will be valuable and necessary for a more comprehensive understanding of motor action planning and execution.

Conclusion
In this study, we explore how humans make realworld decisions by examining the relationship between neural sources of driving behavior and their arousal correlates.We isolate neural ROI that predict steering based on their oscillatory power and found that these regions are functionally organized into distinct networks for scene processing and motor action selection.Additionally, we find correlations between activity in these areas, pupil-linked arousal, and saccades, which highlights the interconnected nature of the brain, body, and behavior for motor decisions.

Methods
First, we describe the paradigm we use to study motor decisions in a realistic VR scenario.Next, we describe the present study on motor action selection, which is done through two steps: (1) predicting motor intensity using oscillatory EEG activity and (2) correlating activity in neural sources with ET and ECG measures during the premotor period.
Participants 12 healthy Columbia University students consented to participate in a part of a larger, three-session study for financial compensation ($20/hour for 3 hours).Ten participants completed all sessions, and only their data were included in the subsequent analyses.All participants provided written consent in a manner approved by the Columbia University Institutional Review Board.They had prior driving experience, normal or corrected-tonormal vision, and reported they were not prone to motion sickness.
At the start of the session, ET calibration is conducted using the VIVE Eye and Facial Tracking SDK from within the experiment.During the experiment, EEG data were recorded at a sampling rate of 2048 Hz with 64 Ag/AgCl pin-type active electrodes, which were connected to a Biosemi ActiveTwo amplifier (Biosemi, Amsterdam, The Netherlands).The electrodes were placed according to the international 10-20 system.All electrode impedances were less than 50 kΩ.ECG data were recorded with two flat-type active electrodes at 2048 Hz by placing one around the sternum and the other around the left clavicle.
Participants used the Logitech G920 steering wheel, which requires continuous force to maintain non-zero steering deviation.

Visual noise and the VR environment
Task difficulty was modulated using the opacity of visual noise presented to participants.This was meant to emulate the amount of fog in driving environments, which affected environmental visibility.In our case, an increase in opacity increases the perceived white-colored 'brightness' by the participant, hence requiring luminance-based adjustment.
We used the Volumetric Fog & Mist bundle [68], a popular asset from the Unity Store, to track the HTC Vive-linked camera object, which participants view in VR, during the BAT.The general idea behind volumetric fog is to create a 3D representation of the effect and render it in a way that realistically interacts with the lighting and other objects in the scene.Our fog object used Unity's directional light shadow map and allowed us to modify the opacity by linking it to voice feedback.An increase in opacity made the surroundings (including the road) more difficult to see and corresponded with an increase in the perceived 'fogginess' in the environment.
To compute differences in connectivity that may be the result of visual noise, we split each trial, at the participant level, across all sessions, by the mean opacity.This resulted in an even split of trials prior to the calculation of PLI.
At the beginning of each block, a path from a random start node to a destination node in a 10×10 grid was selected.The level of fog was determined by a participant's voice-based feedback, which was cued at the start of each vertex in the grid.Participants were instructed to provide voice feedback by speaking either 'easy' or 'hard' to assess the perceived level of difficulty for each trial.The change in opacity indicates the beginning of a new trial.The new trial's difficulty was determined using a staircase design based on the feedback provided by the participants.This approach has been previously used in studies of psychophysics [69,70].
To control for luminance effects on pupil diameter, participants underwent a brief exposure period (∼1 min) before each session and approximately 15 min into the session.They were instructed to remain still and silent while the visual opacity was incrementally adjusted.

Boundary collisions
Because our goal was to study pupil-linked arousal in a driving BAT, we studied differences in connectivity during trials where participants collided with the boundary and when they did not.Each 'trial' in our study corresponds to a single steering action, but multiple turns were required for each portion of the driving path that was dynamically generated per block.Thus, damage from a single collision event is attributed to multiple detected steer events.We did not record collisions for each trial since a single collision may be followed by multiple turns as a participant navigates back toward the main driving path.
Because the number of trials with collision events (983) was much smaller than the number of overall trials (5126), we randomly selected 983 non-collision trials to form a balanced dataset to compute PLI differences between 'Yes' and 'No' collision events in figure 5.

5.2.
Modalities used for characterizing the premotor period 5.2.1.ET Using a moving average filter, we smoothed raw ET data during the premotor period.For each participant, we used a linear fit of pupil diameter vs. opacity during the exposure period to compute their luminance-driven response given a change in opacity.This expected response was subtracted from their phasic pupil diameter trial-by-trial, given the difference in opacity between the last trial's opacity and the current.
The luminance-adjusted pupil diameter, calculated for each premotor period, was used to determine the level of pupil-linked arousal associated with each trial.For all subsequent analyses, we used the left pupil's diameter.
We used the eye tracker's x and y gaze points to classify segments (figure 3(C)).After converting the coordinates, in pixels, to spherical angles in degrees, we used the Naive Segmented Linear Regression method based on Hidden Markov Models [71] for segmentation.The method integrates denoising into segmentation and performs classification on the denoised segments into fixations, saccades, smooth pursuits, and post-saccadic oscillations.It has previously been applied to noisy data to recover gaze position, and velocity estimates in experiments with complex gaze behavior [72,73].Although the method estimates the signal's noise level and determines gaze feature parameters from human classification examples in a data-driven manner, our empirical testing showed that performance was drastically improved if the estimation was performed using at least 3 s of data.Thus, we used 3 s of lookback time relative to the start of each premotor period to fit segments and only extracted saccade-related features during the premotor period.These features included the count and average duration of saccades during the 1-second period.

ECG
In order to characterize the autonomic state, we first used the raw ECG waveform to calculate beats per minute (BPM) using HeartPy [74], a heart rate analysis toolbox validated in human factor research (figure 3(B)).The resulting N-N intervals or normalto-normal peak intervals were used in the calculation of the commonly-used report of HF-HRV, the RMSSD.
Because the premotor period was too short to compute meaningful N-N intervals, we used the 30 seconds leading up to each end of the event time to compute this metric.We only utilized realistic values (50 < BPM < 200) for subsequent analysis.

EEG
Sensor-level EEG data were preprocessed and used to approximate the locations of active sources.A bandpass filter was first applied to the raw, re-referenced EEG data to eliminate frequencies lower than 1 Hz or higher than 55 Hz.Next, the EEG data were downsampled from 2048 to 128 Hz to reduce the time for the remaining processing steps.Subsequently, independent component analysis (ICA) was performed on the EEG data to remove blink-related components through a template-matching technique [75].The template was determined by manually inspecting the ICA components from a random participant to select those that resembled blink-related artifacts, and only the first 30 components, corresponding to the greatest explained variance, were used for subsequent analyses.
After the removal of artifacts using ICA, the data were epoched using the onset of the motor actions, i.e. the interval before each detected motor action.Each epoch has 1 s of data and baseline corrected with 0.2 s prior to the trial onset.
Next, we removed noisy and corrected low-quality epochs using the Autoreject algorithm [76].We estimated the optimal peak-to-peak threshold from a set of candidate thresholds to minimize the error between the mean of training samples and the median of the validation fold across ten folds.Candidate thresholds were determined using Bayesian optimization and applied to the training sets to classify 'good' trials.Using the calculated 'error,' we repaired outlier data segments for each sensor, rejecting ones where the signal was too far above the threshold.The approach has been validated in multiple datasets and inspired by similar, robust, cross-validation methods [76][77][78].
After bad epoch removal and interpolation, we identify sources of EEG activity.Cortical current source density (CSD) was estimated using cortically constrained, low-resolution electrical tomographic analysis (eLORETA).eLORETA is a distributed inverse imaging method where dipolar sources in the cortical areas are estimated jointly.We initially solve the 'forward solution' using the FsAverage MRI template [79] and subsequently obtain an inverse operator.We apply a linear minimum-norm inverse method on motor epochs to compute the source estimates, defined on a source space and average within select ROI on Brodmann's atlas.
Thirteen ROIs were defined a priori from previous literature [15,17,48,49,[80][81][82][83] which have been shown to be relevant in motor behavior under uncertainty in humans.These included areas in the posterior and anterior cingulate cortices (ACC/PCC) as well as bilateral regions in the middle frontal gyrus, supplemental motor area, parietal cortex, motor cortex, and lateral occipital cortex.The five ROIs most relevant to the study after subsequent analysis are highlighted in figure 1(A).

Statistical and connectivity approach
We first remove the effects of luminance-induced pupil diameter changes through a simple linear regression analysis [84] from the opacity exposure period.For subsequent analysis, we assume that the pupil-diameter changes correspond to changes in pupil-linked arousal.
Prediction of absolute steering deviation We analyzed the effects of cortical ROI by frequency band on motor actions.Steering deviation was predicted using a LMEs model.Mixed effects modeling is commonly adopted in neuroscience studies [85,86] and was chosen since it holds several advantages over traditional F-tests in assessing predictive actions.These include the ability to account for individual differences in participant characteristics, which was achieved in the present study by modeling each participant's intercept as a random effect.The magnitude of detected steering wheel actions was used as the dependent variable.Band-and ROI-specific spectral power were modeled as fixed effects.We isolate ROIs which are significant predictors of steering deviation, and conduct further study.
To investigate the role of significant ROIs in decoding motor activity, we fit a linear regression model using single-ROI source activity at each timepoint as regressors and evaluated its performance on held-out motor epochs.This approach estimates discriminative spatial filters and is similar to sliding estimator-based approaches in fMRI.This method allows for the prediction of experimental outcomes and interpretation of when the effect of interest occurs [87].In our study, we assessed the performance of individual ROIs using the variance explained (R 2 ) measure and averaged the activity from 100 folds.
Connectivity we attempt to validate the initial hypothesis of two distinct functional connectivity networks, namely 'scene processing' and 'motor action selection' networks.We compute phase lag index [44], a commonly used measure of statistical interdependencies for source time courses between our isolated ROIs' time series.We selected the phase lag index over other functional connectivity measures due to its reported insensitivity to the effects of volume conduction [88].We performed this analysis for frequencies ranging from 4-55 Hz using a Hanning multitaper time-frequency transformation of the epoched trials between −1 to 0 s, locked to the motor action onset to capture the planning period.
We assessed whether the observed PLI values are statistically significant through a permutation-based approach [44].Generally, in figure 5 (A), the comparison is between surrogate data, after randomly shuffling the phase values of each source time series, creating a distribution of chance PLI values, while keeping the other signal intact.We use this distribution of surrogate PLI values for determining the significance of our observed PLI.In figure 5(B), the significance is between experimental conditions.
Correlations between modalities lastly, we attempt to validate that the 'motor action selection' network is more strongly related to pupillinked arousal state, autonomic state (through HRV), and saccade behavior (count and duration) than the 'scene processing' network during the premotor period.We conduct Spearman correlation analyses between band power in significant ROIs and measures from each modality.Specifically, we target RMSSD as a measure of HF-HRV, saccade duration as a measure of cognitively driven looking behavior, and left pupil diameter as a measure of pupil-linked arousal.
In summary, our study explores how humans make real-world decisions by examining the relationship between neural sources of driving behavior and their arousal correlates.We isolated neural ROI that predict steering based on their oscillatory power and found that these regions are functionally organized into distinct networks for scene processing and motor action selection.Additionally, we observed correlations between activity in these areas, pupil-linked arousal, and saccades, which highlights the interconnected nature of the brain, body, and behavior for motor decisions.

Figure 2 .
Figure2.Overview of how motor events were detected and premotor period is defined.(A)We simultaneously collected neural data from EEG, autonomic measures using ECG, eye movements and pupil dynamics using a VR-headset embedded eye tracking system, and motor actions using a steering wheel.(B) Participants (n = 10) performed three virtual reality driving task sessions, requiring boundary avoidance under time pressure and changing visual uncertainty.Their motor actions were recorded from the steering deviation as they were navigating a city environment.We analyzed direction-independent (i.e.absolute) steering deviation).Motor actions belong to a global trial with a set level of visual fog (opacity) in the environment that participants drove in.(C) The start of each motor action was marked using a peak detection method on the steering wheel data since this was most relevant to navigating the BAT.The premotor periods of interest were a fixed, 1-second interval before each event, and the intensity of the motor activity was determined by the post-event steer angle.Blue and red circles indicate events with low and high motor intensity, respectively.

Figure 3 .
Figure 3. (A)Event locking was done in reference to peaks and troughs in the steering wheel channel.This determined the end of each 1-second event.Specifically, steer, throttle, and break events were acquired as the degrees of deviation from the wheel's center position.(B) Electrocardiogram (ECG) acquired in amplitudes (mV).(C) Fixation, saccade, smooth pursuit as well as post saccadic oscillations (PSO) acquired as degrees of horizontal gaze position.(D) A single session's motor actions, from the peak (red, corresponding to a right turn) or trough (green, left turn) identified in(A).Whereas the actions in (A) were used for event locking, the intensity of motor actions attributed to each event was binned using the absolute value of the steering deviation (i.e.'final' steering wheel degree), with respect to the start of the event, after 750 msecs.(E) Distributions, by-participant, of the primary outcome of interest, the absolute steering deviation, characterizing the magnitude of motor decisions.

Figure 4 .
Figure 4. (A) Raw pupil diameter from the left pupil and linear regression fits.The effect of trial opacity on left pupil diameter was residualized from subsequent analysis to capture non-luminance-driven, pupil-linked arousal.Each subplot shows raw pupil diameter and fits across five opacity levels used during the exposure period for a participant.(B) Absolute steering deviation (in degrees) vs. luminance-adjusted pupil diameter (dotted reference line at 3 mm).(C) Sample session from a single participant.Steering action onset (relative to the start of the session) vs. steering deviation is displayed.The size of the points indicates the pupil diameter averaged during the premotor period associated with each steer event; the color indicates the opacity of the visual noise (i.e.fog) during the trial when the steering action occurred.The participant took a rest from approximately 620 s into the trial until 780 s.

Figure 5 .
Figure 5. (A) Phase lag index (PLI) connectivity measures between significant Regions of Interests (ROIs) relevant for prediction of absolute steering deviation across all motor epochs, per-band (θ: 4-8 Hz, α: 8-15 Hz, β: 15-32 Hz, γ: 32-55 Hz).ROIs include visual cortex areas (V1, V2), ventral (vACC) and dorsal (dACC) anterior cingulate cortices, and dorsolateral prefrontal cortex (DLPFC).LH and RH refer to the left and right hemispheres, respectively.Significance is calculated using random phase-adjusted data.((B), top) Differences in PLI between high and low visual opacity.The mean trial opacity was used to split trials, by participant across all sessions, into 'high' and 'low' opacity conditions.((B), bottom): differences in PLI between trials where participants collided with the road boundary ('yes' collision) and did not ('no' collision).Because only 19.2% of trials had a collision event detected, we randomly selected 983 matching trials with no collision events prior to computing differences in connectivity of the balanced dataset.* * * = 0.001, Mann-Whitney U test, followed by false-discovery rate-based correction for multiple comparisons.

Figure 6 .
Figure 6.Correlated activations: voxel-wise Spearman correlation coefficients for relevant measures (column) by hemisphere (rows) on medial views of the brain.LH and RH refer to the left and right hemispheres, respectively.The colorbar represents ρ, the correlation coefficient calculated independently for each measure.ROIs, identified as significant for predicting motor action in our study, are demarcated using labels from figure1(A).RMSSD, root mean square of successive differences.