time depletes resources of top-down control:
EEG evidence
Abstract
Mental fatigue is detrimental to our ability to remain attentive and vigilant, though the exact nature of this effect is still very much unclear. Decreases in performance after extended participation in a vigilance task (known as the time-on-task effect) have been explained in terms of depletion of cognitive resources, but no substrates representing these resources have been identified to date. We used electroencephalography (EEG) to examine whether the time-on-task effect is associated with changes in top-down attentional re-sources. Prolonged time-on-task is expected to result in a shift away from top-down control of attention to a more stimulus-driven strategy, which is manifested in the brain as a transition to a different mode of atten-tional processing. To induce mental fatigue, participants performed a sustained attention task for 80 minutes without breaks. Our results show that perceptual sensitivity (𝐴′ scores) indeed declined over time, which
could not be solely attributed to demotivation. Stability of attentional processing closely mirrored this trend in behavioral performance, as indexed by a decreased trial-to-trial consistency of stimulus-locked phase values in the theta band over right parieto-occipital (rPOC) cortex and mid-frontal (MFC) electrode sites. These dy-namics appeared to be accompanied by diminished functional connectivity of rPOC and MFC. Lastly, we identified tonic changes in alpha- and delta-band power that were coupled to target perception and theta phase coherence, respectively. These findings suggest that sustained attention suffers from depletion of limited top-down resources and thereby provide novel fundamental insights into the short-term plasticity of attentional processes.
Keywords: sustained attention, fatigue, resource depletion, EEG, theta (4-7 Hz), alpha (8-12 Hz), delta (1-3 Hz).
INTERNSHIP REPORT
Author: Leon Reteig (Student ID: 5826705) Date: 12-11-2012
Program: research Master Brain & Cognitive Sciences, Cognitive Neuroscience Track Supervisor: dr. Heleen A. Slagter
Introduction
Mental fatigue following demanding cognitive exer-cise is a common experience for many people. Con-tinuous mental effort often leads to decreases in performance over time, even when the potential consequences are disastrous, such as for air force personnel (Bartlett, 1943; Mackworth, 1948). This phenomenon of declining performance over periods of sustained attention is known as the time-on-task effect (Ackerman, 2011), also sometimes called the vigilance decrement (Helton & Warm, 2008). How-ever, despite its ubiquity in daily life and in academic research, the underlying causes of the time-on-task effect remain largely unknown.
One prominent theory frames the time-on-task effect in terms of cognitive resources (Helton &
Warm, 2008). In this view, successful monitoring of
incoming information requires dedication of limited attentional resources to the task at hand. When maintaining vigilance over lengthy periods of time, the pool of resources is depleted, thereby progres-sively worsening performance. A growing body of literature provides empirical support for the resource model (see Warm et al. (2008) for a review). These studies show that performance is more greatly af-fected when task difficulty is high (Smit et al., 2004)
and/or signal salience is low (Helton & Warm, 2008). Correspondingly, participants report sustained atten-tion tasks to be effortful and demanding (Grier et al., 2003; Szalma et al., 2004).
However, resource depletion is not necessarily the only factor contributing to the time-on-task effect. Others have proposed that the performance decre-ment is due to mindlessness, manifested by task disengagement following boredom and increased presence of task-unrelated thoughts (Manly et al.,
1999; Robertson et al., 1997). A related approach
notes that since attentional focus is energetically demanding, one would only be able to sustain it in case of sufficient motivation and little aversion
to-wards the task (Hockey, 1997), i.e. provided that the benefits outweigh the costs (Boksem & Tops, 2008).
Ultimately, further fundamental insights into the time-on-task effect and its contributing factors could best be garnered by also examining their neurophys-iological underpinnings. Yet, the time-on-task effect has received surprisingly little attention from cogni-tive neuroscience. Boksem et al. (2005) conducted an electroencephalography (EEG) study and found increases in theta and alpha band power with time-on-task. Similarly, mental fatigue has been associat-ed with widespread changes in EEG power and coherence (Lorist et al., 2009). These results were not coupled to specific task parameters or cognitive processes, though two event-related potential (ERP) studies by the same group suggest that top-down control of attention is specifically affected (Boksem et al., 2005; Lorist, 2008).
The aim of the current study was to investigate the psychological and neural correlates of the time-on-task effect. Specifically, we ask whether electro-physiological indices of top-down attentional control are modulated by time-on-task, and how such changes relate to concurrent shifts in behavioral performance. To this end, participants performed a spatial attention task involving visual discrimination of rare targets from standard non-targets presented in the left visual hemifield. They did so for 80 minutes without breaks while EEG was measured. After 60 minutes, participants were motivated to increase their performance through a social and financial incentive. We expect to find the classic vigilance decrement effect, i.e. that perceptual sensi-tivity will decrease over time. If this decrement is indeed caused by depletion of cognitive resources, the motivation manipulation should not be sufficient to fully rescue performance.
We hypothesize that the time-on-task effect in-volves a transition from proactive control to a reac-tive control-based strategy (Braver et al., 2007). Proactive control is a top-down, preparatory manner of orienting attention that is highly effective but
en-ergetically costly. Proactive control is mediated by a rhythmic mode of processing in which neural activity is phase-locked to task events (Schroeder &
Lakatos, 2009). Once top-down attentional
re-sources are depleted, participants are forced to switch to a reactive control scheme instead. This type of cognitive control is largely stimulus-driven, which is less demanding but also suboptimal, partic-ularly in conditions where stimulus information is weak (Braver et al., 2007).
This attenuation of top-down control should be reflected in the EEG results in three different ways. First, as time-on-task increases, the readiness of the attentional system to process new information likely declines. As attentional stability decays, the re-sponse to incoming stimuli should grow more varia-ble. Previous work of ours shows that mental train-ing-related improvements in target perception and sustained attention are accompanied by an in-creased coherence of the phase of oscillations in the theta (4-7 Hz) band (Lutz et al., 2009). Phase co-herence is a measure of the cross-trial variability of stimulus-evoked neural activity. Presently, we there-fore predict the reverse effect: since the breakdown of attentional stability should lead to more variability in stimulus-locked responses, theta phase coher-ence will decrease as a function of time-on-task.
Second, attention is mediated by a wide variety of brain areas, particularly in the frontal and parietal lobes (Corbetta & Shulman, 2002). Since top-down attentional control depends on the functional integri-ty of these networks, the time-on-task effect likely involves weakened neural communication. Changes in long-range communication between neural sites can successfully be revealed by examining phase synchronization (Cohen & van Gaal, 2012; Palva &
Palva, 2011). Aside from the fronto-parietal attention
network, the medial prefrontal cortex is often re-garded as a key area for cognitive control (Cohen et
al., 2009; Ridderinkhof et al., 2004). We therefore
predict that prolonged time-on-task produces a
de-creased functional connectivity of these areas, par-ticularly in the theta and alpha frequency bands.
Third, top-down allocation of spatial attention, i.e. the ability to maintain focus on the area in which target stimuli can appear, may also be impaired in the time-on-task effect. Recent studies suggest that alpha (8-12 Hz) oscillations play a major role in inhibitory control in general, and spatial attention in particular (for reviews, see Jensen and Mazaheri
(2010); Klimesch et al. (2007)). Alpha rhythms over
parietal and occipital cortices appear to exhibit a spatial attention bias: alpha power is suppressed in the (contralateral) hemisphere that processes visual information, whereas it is increased in the ipsilateral hemisphere (Sauseng et al., 2005). These alpha dynamics are anticipatory and predict perception of the upcoming visual stimulus (Mazaheri et al., 2009;
Romei et al., 2010; Thut et al., 2006). We expect
that a shift away from top-down control negatively affects the orienting of spatial attention, which should be evidenced by a disintegration of the alpha power asymmetry with time-on-task.
Lastly, we also conducted an exploratory analy-sis of slow (0.1 – 5 Hz) oscillations, since they have been recently implied in the ability to perceive weak signals (Monto et al., 2008).
Materials & Methods
ParticipantsThirty subjects were recruited to participate in this experi-ment from the University of Amsterdam student popula-tion. Nine participants were excluded from further analysis due to not having completed the task (2 subjects), poor EEG data quality and/or insufficient correct responses. All data presented here, including behavioral data, are from the remaining 21 participants (11 female; mean age: 21.6, SD: 2.3). All participants had normal or corrected-to-normal vision, no history of mental or neurological disorders and were excluded from participation if they reported getting more than two hours less sleep than usual the night prior to the experiment. The study was approved by the local ethical committee. All
partic-ipants gave their informed consent and were paid the nominal fee of € 7,- per hour, with a chance to receive an extra amount based on performance (see Procedure). Task
This study used a modified version of the sustained atten-tion paradigm employed by MacLean et al. (2009;
experiment 2, transient version). In this task, participants
are required to visually discriminate briefly presented rare target stimuli (short lines) from standard non-targets (long lines). Participants were instructed to respond with a but-ton press when viewing a short line, but to withhold re-sponses for long lines.
Participants sat at a viewing distance of 90 cm in front of a 17-inch BenQ TFT monitor with a refresh rate of 60 Hz. Light gray lines were presented on a black back-ground 3 degrees to the left of a yellow fixation point. The lines remained on screen for 150 ms, accompanied by a mask pattern presented for 100 ms before and after each line stimulus (Figure 1). Line stimuli were presented with an inter-stimulus interval of 1850 ms, resulting in a high event rate of 30 events/min to ensure sufficient task de-mand (Parasuraman, 1979). Participants were instructed to maintain fixation on the yellow dot at all times, thus having to covertly attend to the area in which targets were presented. The experiment was set up in this manner to avoid excessive eye movements which are picked up by the EEG equipment, and because our hypothesis on spa-tial attention required that all visual responses were con-fined to one hemisphere only.
Figure 1 | Sustained attention task. Participants judge the length of the line stimulus presented for 150 ms while fixating their eyes upon the yellow dot at all times. Their assignment was to respond only when a rare target (short line) was presented and not when viewing a more frequent non-target (long line) A target trial is shown; stimuli and timing parameters for non-target trials are identical except for line length.
Prior to the start of the main task, individual task dif-ficulty was calibrated for each participant using Parameter Estimation by Sequential Testing (PEST) (Taylor &
Creelman, 1967; cf. Maclean et al. (2009)). PEST is a
thresholding procedure that adaptively changes the step size between testing levels to estimate the desired level of an independent variable. In the present study, the proce-dure adjusts the length of the short line until a stable per-formance of 80% accuracy is reached. The only difference between the PEST procedure and the main task was a higher target to non-target ratio (1:3.5 vs. 1:5).
Long lines were always 4.5° in length, whereas line width (0.035°) was constant for both long and short lines. The mask was composed of many short lines (0.07° × 0.28°-0.45°) positioned within a 1.0° × 5.0° space. To prevent participants from assessing the line length of stimuli by comparison to the length of the lines comprising the mask, each mask element was vertically repositioned by a random amount (within -0.14° to +0.14°) upon each presentation.
Procedure
Following an on-screen tutorial of the task instructions, the thresholding procedure was first trained for one minute. Subsequently, participants performed the full PEST pro-cedure. Because of the adaptive nature of PEST, the procedure has no fixed length, though PEST duration for all participants varied only modestly (between 7 and 13 minutes). During the PEST procedure, participants were given auditory feedback indicating a hit (ding sound), a miss (woosh sound) and a false alarm (also a woosh
sound). After PEST, the main task was trained for one minute, during which participants still received feedback in order to let them adjust to the calibrated length of the target stimuli. Finally, the main task was performed for 80 minutes without breaks in order to keep participants from replenishing their attentional resources (Chen et al.,
2010). Each participant completed 2400 trials in total, or
300 trials per 10 minutes. The ratio of target to non-target stimuli was 1:5, resulting in 60 target trials per 10 minutes.
Every 10 minutes, participants were asked to rate both their motivation to perform well and their aversion towards the task on a seven-point scale (1: no aver-sion/motivation; 7: strong aversion/motivation). They had 6 seconds to do so by moving a cursor to the number best reflecting their current level of aversion/motivation. Fixation dot Mask Mask Target 1650 ms 100 ms 100 ms 2000 ms 150 ms
After performing the task for 60 minutes, a new screen was displayed informing participants of a chance to gain an additional sum of money – an option that was unknown to them up until then. This manipulation was designed to motivate participants to increase their perfor-mance, in order to evaluate whether motivation can bring attentional resources back online. They would receive € 35,- on top of their nominal participation fees if they out-performed at least 65% of the other participants during the last 20 minutes of the task (cf. Lorist et al., 2009). This information was presented until the participant pressed a button indicating he/she had understood it. All participants pressed within one minute, so none of the participants had any substantially larger amount of rest than the others. Participants again reported their subjective motivation and aversion immediately before and after the motivation in-struction.
Behavioral analyses
Our main index of behavioral performance was expressed as 𝐴′, a nonparametric measure of perceptual sensitivity
from signal detection theory (Stanislaw & Todorov, 1999).
A’ is dependent on hits, misses, false alarms (FAs) and correct rejections (CRs), and is calculated as follows:
𝐴′= .5 +(𝐻 − 𝐹)(1 + 𝐻 − 𝐹)
4𝐻(1 − 𝐹)
If 𝐻 > 𝐹, where 𝐻 is Hit rate [Hits / (Hits + Misses)] and 𝐹 is false alarm rate [FAs / ( FAs + CRs)]. 𝐴′ can take any
value between 0.5, meaning that target stimuli are indis-tinguishable from non-targets, and 1, signaling perfect performance. 𝐴′ values were computed separately for
every 10 minutes of the experiment.
EEG data acquisition and preprocessing EEG data were recorded at 512 Hz using a BioSemi Ac-tiveTwo 64 Ag-AgCl channel setup (BioSemi, Amsterdam, The Netherlands) placed according to the international 10-20 system. Four external electrodes recorded the electro-oculogram from vertical and horizontal ocular sites; two additional electrodes were placed on both earlobes. EEG data were high-pass filtered offline at 0.1 Hz and re-referenced to the average of the earlobe electrodes.
The continuous EEG data were then segmented into epochs from -2000 ms to +2000 ms peri-stimulus, of which 1 second buffer zones on each end were meant to ac-commodate the edge artifacts that may result from wavelet convolution (see Time-frequency decomposition). All
trials were visually inspected; those containing artifacts due to electromyographic (EMG) activity or other sources of noise were removed. Independent component analysis was performed using the EEGLAB toolbox (Delorme &
Makeig, 2004) operating in the MATLAB (Mathworks)
programming environment. Components containing eye blink, oculomotor or other artifacts clearly distinguishable from genuine neural activity were subtracted from the data.
Epochs were separated into different conditions ac-cording to trial-type (hits, misses and CRs). In order to track changes in neural activity over time, epochs were further partitioned into eight 10-minute blocks for CRs and four 20-minute blocks for hits and misses. Because each block had relatively few target trials compared to non-target trials, there were insufficient amounts of hit and miss trials to warrant splitting them into the finer-grained 10-minute blocks as well. Trial counts were equalized over blocks per participant, such that the same number of CR trials per 10 minute block (M: 171.3, SD: 25.3) and the same number of hit and miss trials per 20 minute block (M: 23.8, SD: 5.4 ) were used in analyses. FA trials were too few for EEG data analyses, even for 20 minute blocks (M: 9.7).
Time-frequency decomposition
Time-frequency representations of the EEG data were obtained using custom scripts written in MATLAB supple-mented by code from the EEGLAB toolbox (Delorme &
Makeig, 2004). All epochs were concatenated into one
long timeseries and subsequently convolved with a family of complex Morlet wavelets (cf. Cavanagh et al., 2009). These wavelets consist of a complex sine wave tapered with a Gaussian window:
𝑒−2𝑖𝜋𝑓𝑡 ∙ 𝑒−𝑡2𝑠22
where 𝑓 is frequency, 𝑡 is time, and 𝑠represents the width of the Gaussian. Frequency increased from 2 to 80 Hz in 30 logarithmically spaced steps. 𝑠 equals 𝑥/2𝜋𝑓, where 𝑥 increased logarithmically from 3 to 12 in the same number of steps. This dynamic increase of the number of wavelet cycles was employed in order to obtain a better frequency resolution at the expense of temporal resolution as fre-quencies go up, which attenuates the problem that high-frequency dynamics exhibit very little phase-locking. Fol-lowing convolution, data were reshaped back into individ-ual epochs. Concatenation and subsequent reshaping was
performed primarily for computational efficiency and also to minimize edge artifacts. Edge artifacts were not re-moved but were instead confined to 1s long buffer zones at both extremes of each epoch. Cutting the epochs this way during preprocessing and refraining from analysis of these buffer zones is a convenient way to deal with edge artifacts without transforming the data.
In addition to this general analysis, we also ran a separate wavelet convolution tailored to expose time-frequency dynamics of very slow oscillations. Here, fre-quency increased from 0.1 to 5 Hz in 11 logarithmically spaced steps, and frequency band width was held con-stant at 3/2𝜋𝑓. To accommodate the larger edge artifacts at low frequencies, each epoch was mirrored and spliced to itself at both ends, yielding 15s epochs with larger edge artifact buffers.
The result of wavelet convolution is a complex num-ber 𝑍, which can be construed as a point in the complex plane. From this time-frequency representation of the data, we extracted two orthogonal measures of oscillatory dynamics: trial-averaged power and phase values.
Power 𝑝 at time point 𝑡 represents the amount of en-ergy in the signal at a particular frequency band, and is defined as the modulus of the vector pointing from the origin to 𝑍: 𝑝𝑡= 𝑟𝑒𝑎𝑙(𝑧𝑡)2+ 𝑖𝑚𝑎𝑔(𝑧𝑡)2. Power values were
baseline corrected to the average power from 1000 to -800 ms and converted to a logarithmic scale by a decibel transformation: 𝑝𝑜𝑤𝑒𝑟𝑑𝐵= 10 ∙ log10[𝑝𝑡/𝑝𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒]. Such
logarithmic scaling allows one to make valid comparisons across different frequency bands, because it corrects for the ‘power law’ stating that power at low frequencies is intrinsically higher. All 10 minute blocks (CR trials) shared an average baseline, as did all 20 minute blocks (hit and miss trials), such that we could make comparisons be-tween different blocks of CR trials, as well as bebe-tween hit and miss trials. Baseline times were chosen to minimize interference from both anticipatory pre-stimulus activity and post-stimulus processing.
Phase 𝜑 at time-frequency point 𝑡 is defined as the angle of aforementioned vector to the origin: 𝜑𝑡=
arctan [𝑖𝑚𝑎𝑔(𝑧𝑡) / 𝑟𝑒𝑎𝑙(𝑧𝑡)] . Two different types of
oscilla-tory phase measures were computed: inter-trial phase coherence (ITPC) and inter-site phase synchronization (ISPS). ITPC reflects the consistency of phase values over trials at a particular electrode, and is defined as:
𝐼𝑇𝑃𝐶 = �1𝑛 ∙ � 𝑒𝑖𝜙𝑘(𝑡)
𝑛 𝑘=1
�
where 𝑛is the number of trials and 𝑒𝑖𝜙𝑘(𝑡) is the complex
representation of the phase angle 𝜙 on trial k, at time-frequency point 𝑡. ITPC can be interpreted as a measure of trial-to-trial stability of evoked brain responses.
In contrast, ISPS measures the extent to which phase angle differences between two electrodes are con-sistent over time and frequency:
𝐼𝑆𝑃𝑆 = �1𝑛 ∙ � 𝑒𝑖[𝜙𝑘𝑗(𝑡)−𝜙𝑘ℎ(𝑡)]
𝑛 𝑘=1
�
where 𝑗and ℎ are different electrodes. ISPS can be inter-preted as a measure of functional connectivity between different neural sites. A potential problem for interpretation of ISPS values is volume conduction, i.e. the spread of electrical activity throughout the head due to the conduc-tivity of brain tissue and the high resistance of the skull. Volume conduction can cause neural activity from a single source to contribute to different electrode sites over the scalp, which is hard to distinguish from genuine synchro-nization between multiple neural sources. However, vol-ume-conducted activity can be minimized by applying the Current Source Density (CSD) -transform (Tenke & Kayser, 2005). CSD is a spatial filter that increases spatial selectivity by highlighting local dynamics, and has been validated for analyses of ISPS (Srinivasan et al., 2007).
Both ISPS and ITPC vary between 0 and 1, where 0 means no phase consistency and 1 means a perfect con-sistency of phase values. Note that for ISPS, the phase values do not have to be equal; there can be a phase lag between two electrodes, as long as the latency is con-sistent over time and frequency. Absolute ITPC values are reported, since there is virtually no spontaneous, ongoing phase coherence in the absence of external events. In contrast, to account for small differences in baseline ISPS between blocks, all values are expressed as a percentage change from baseline (-1000 to -800 ms). For the sake of convenience, ITPC will be referred to as phase coherence, whereas ISPS is denoted as phase synchronization. Statistical analyses
Changes in 𝐴′ and neural activation over time were
as-sessed through repeated-measures ANOVAs. Green-house-Geisser corrected p-values are reported in case of sphericity violations; α was set to 0.05. The within-subjects
factor TIME consists of either 8 levels (for 𝐴′ values and
CR trials, split into 10 minute bins) or 4 levels (for hits and misses, split into 20 minute bins). To test for significant differences between time blocks, we used a simple gen-eral linear model (GLM) contrast in which the mean of each block is compared to the mean of the first block (i.e.
the first 10 minutes).
In repeated measures designs, graphs showing typ-ical standard errors are not very useful, since these incor-porate the often very large between-subject variability that obscures consistent within-subject effects. Therefore, we transformed data for plotting purposes with error bars in the following manner: 𝑌 = 𝑋𝑖𝑗− 𝑋� + 𝑋�, where Y is a par-𝚥
ticipant × condition data matrix, 𝑋𝑖𝑗 is one element of Y at
condition 𝑖 and participant 𝑗, 𝑋� is the subject mean of 𝚥
participant 𝑗 and 𝑋� is the group mean (Cousineau, 2005). This transformation preserves the condition means of the original data, but all subject means are now identical to each other and to the group mean. Lastly, variances are corrected by a scaling factor 𝑀/(𝑀 − 1), where 𝑀 is the number of within-subject conditions (Morey, 2008).
Repeated measures ANOVAs for EEG data were computed on the basis of specific windows in time-frequency space. These windows were selected based on condition-average time-frequency maps over all subjects and are thus orthogonal to differences between particular conditions. We used two different methods to extract data from these windows for statistical analyses. For more stable effects lasting several 100 ms, all values in the condition-average window were averaged together for every condition, to obtain one value per subject, per condi-tion. For short responses with clearly-defined peaks, one peak value per subject was extracted by (1) finding the pixel with the highest absolute value within the condition-average window, (2) drawing a smaller subject-specific window around it and (3) for every condition, averaging across all pixels within this new window. This latter ap-proach was favored over just averaging over all values in the condition-average window to accommodate small inter-individual differences in peak latency and frequency bands.
To examine whether changes in behavioral perfor-mance were mirrored by concurrent changes in neural activity, we correlated the perceptual sensitivity curve over all blocks with the extracted EEG measures. Computing correlations between two variables with repeated
observa-tions is a highly non-trivial problem, since it is not evident how best to model the within- and between- participant variance. In these cases, we chose to compute individual Spearman correlations for each subject, yielding as many correlation coefficients and p-values as there are subjects.
r-values were transformed to Z-scores using Fisher’s transformation (Fisher, 1915):
𝑍 = 12 ln�1 + 𝑟1 − 𝑟�
From the inverse of this function, a Fisher-weighted mean correlation coefficient 𝑟̅ can then be obtained:
𝑟̅ = 𝑒𝑒𝑧̅𝑧̅− 𝑒+ 𝑒−𝑧̅−𝑧̅
where 𝑧̅ is the arithmetic mean of the individual Z-scores. P-values were combined using Fisher’s method (Fisher, 1925):
𝑋2= −2 � ln(𝑝
𝑖) 𝑘
𝑖=1
where 𝑝𝑖 is the p-value for the 𝑖-th test out of 𝑘 total tests.
When the null hypotheses are true, 𝑋2 follows a
chi-squared distribution with 2𝑘 degrees of freedom, from which an overall p-value can be extracted. We used a Bonferroni-correction of α/𝑘 to correct for multiple depend-ent comparisons. All other correlations are Pearson corre-lations of subject means that thus do not exhibit a repeat-ed-measures structure.
Results
BehaviorPerformance on the sustained attention task changed significantly over time [main effect of TIME, F(2.50, 49.97) = 7.47, p = 0.001] (Figure 2A). In accord with the classical vigilance decrement, per-ceptual sensitivity declined with time on task, as performance peaked within the first 10 minutes (M: 0.91). At the same time, a short but distinct boost in performance is visible in block 7, immediately after the motivation instruction. However, performance in block 7 does not reach the same level as in block 1 (simple GLM contrast, p = 0.026; level 1 vs. all other blocks p’s < 0.001) , and it immediately drops back
down to the lowest level overall in block 8 (M: 0.849).
Figure 2 | Perceptual sensitivity ( ′) and subjective ratings over time. Dashed line depicts moment at which the motivation screen was presented. Asterisks demonstrate significant differ-ences at p < 0.05 compared to the first block/sample (A) ′
val-ues per 10 minutes of the experiment over all subjects. Notice the progressive decrease in perceptual sensitivity with time-on-task, which is temporarily reversed by the motivation manipulation, only to fall back down again in the next block. (B) Motivation and aversion ratings at 10 different sample times. “begin” = before start of the experiment; “10”-“70” = minutes after start of experi-ment; “pre” = just before motivation manipulation (60 minutes after “begin”); “post” = right after motivation manipulation (one minute after “pre”); “end” = after last block of experiment (80 minutes after “begin”). Motivation declines over time, whereas aversion rises. The motivation instruction affected both measures in the opposite direction, such that the final three motivation ratings were not significantly different from the first.
Motivation ratings also varied significantly with time on task [main effect of TIME, F(3.90, 77.80) = 7.10, p < 0.001], as did the aversion ratings [main effect of TIME, F(3.46, 69.17) = 12.67, p < 0.001]
(Figure 2B). Ratings before the task began were
fairly positive, indicating that participants were moti-vated to do well (M: 5.05) and did not harbor much aversion towards the task (M: 2.62). Yet, motivation levels declined as a function of time on task, while the aversion ratings rose over time. Aversion ratings at the start of the experiment differed significantly from all others (simple GLM contrast, p’s < 0.019), whereas the first motivation rating differed from those between 30-50 minutes and from pre-motivation (60 minutes into the task) (simple GLM contrast , p’s < 0.019).
A paired-sample t-test revealed that the motiva-tion manipulamotiva-tion had a significant effect on both ratings: motivation ratings immediately before the instruction were lower than right after its delivery [t(20) = -5.14, p < 0.001]; reversely, the motivation instruction brought about a decrease in the aversion ratings [t(20) = 3.16, p = 0.005]. Importantly, report-ed motivation during the last two blocks of the task and after completion of the task did not differ signifi-cantly from motivation during the first 10 minutes (simple GLM contrast, p’s > 0.084) indicating that the financial incentive successfully motivated people to do well at the end of the task. Thus, despite sub-jects’ high motivation even during the last 10 minutes of the task, their performance dropped to the lowest level. This finding indicates that the ob-served decrease in perceptual sensitivity in block 8 cannot be solely attributed to decreasing motivation, which provides some indirect evidence that the per-formance decrement may stem from depletion of cognitive resources instead.
In addition, we performed the following control analysis that further corroborates the resource mod-el. If motivation were the main contributor to the initial performance decrement, one might expect those participants who exhibited the largest decline
A
B
1 2 3 4 5 6 7 8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 Perceptual sensitivity Block (10 minutes) A’begin 10 20 30 40 50 pre post 70 end
1 2 3 4 5 6
7 Motivation & aversion ratings
Sample
Ratings (1 = none; 7 = fully)
motivation aversion
in 𝐴′ (due to a lack of motivation) to best be able to
rescue their performance after the motivation in-struction, since they would have a greater possibility to enhance their performance. However, we found that the difference in 𝐴′ between block 1 and block 6
(pre-motivation instruction) is uncorrelated to the performance difference between block 6 and 7 (pre- and post-motivation instruction) (Spearman’s rank-order correlation, r = 0.18, p = 0.437).
EEG data
Attentional stability: Effects of time-on task on inter-trial phase coherence
We first hypothesized that attentional stability deteri-orates with time-on-task. We predicted that this pro-cess would be reflected by a decreased phase co-herence of oscillations in the theta band, since theta phase coherence has been positively associated with sustained attention (Lutz et al., 2009).
To this end, condition-average topographical maps of phase coherence during CR trials were examined to precisely locate electrodes and fre-quency bands of interest. As Figure 3A shows, post-stimulus phase coherence in the theta frequency band (4-7 Hz) between 150-450 ms was maximal over right parieto-occipital (Figure 3C1) and mid-frontal areas (Figure 3C2). This effect was accom-panied by a concurrent increase in theta power at the same scalp locations and within the same time-frequency window (Figure 3B). Posterior electrodes that loaded maximally on these effects were PO8, P6 and P8; the anterior site of activation corre-sponded best to FCz, F1 and F2. For subsequent analyses, we pooled data from these three elec-trodes at posterior and anterior sites, which will henceforth be referred to as right parieto-occipital cortex (rPOC) and mid-frontal cortex (MFC), respec-tively. Averaging over multiple electrodes has the added advantage of increasing signal-to-noise ratio and takes into account small differences between subjects in the precise location of the effect.
In line with our predictions, theta phase coher-ence changed as a function of time on task, in both rPOC (Figure 3D1) [main effect of TIME, F(7,140) = 3.17, p = 0.004] and MFC (Figure 3D2) [main effect of TIME, F(3.27,65.38) = 4.75, p = 0.004). Theta power also displayed a significant main effect of time in rPOC [F(3.17, 63.30) = 2.97, p = 0.036], but not in MFC (p = 0.568). Closer inspection of the theta phase coherence effect revealed a distinct resem-blance to the change in behavioral performance over time (Figure 3E, cf. Figure 2A). Similar to 𝐴′,
post-stimulus theta phase coherence exhibited a general negative trend over time, which was briefly changed in polarity after the motivation manipulation. Corre-spondingly, we found that theta phase coherence in block 1 differed significantly from all other blocks (simple GLM contrast, POC: p’s < 0.023; MFC: p’s < 0.016) except block 7 (POC: p = 0.090; MFC: p = 0.334). Correlational analyses (see Statistical
anal-) confirmed a significant positive relation be-yses
tween 𝐴′ and theta phase coherence for both POC
(𝑟̅ = 0.33, p < 0.001) and MFC (𝑟̅ = 0.44, p = 0.002). The progression of the theta power effect was erratic and did not parallel the 𝐴′ curve, so we did not
ex-plore it much further. The presence of theta power is nonetheless informative in itself, since it strongly suggests that the concomitant theta phase coher-ence effect was due to stimulus-evoked activity and not resetting of the phase of ongoing oscillations
(Sauseng & Klimesch, 2008).
The theta phase coherence results indicate that the motivation manipulation initially led to a success-ful recovery of attentional stability and a correspond-ing increase in perceptual sensitivity. However, this boost was apparently unsustainable, as within a mere 10 minutes, both attentional stability and be-havioral performance plummeted back down. In addition, further corroborating that theta phase lock-ing was tightly coupled to successful stimulus per-ception, phase coherence was stronger in hit than in miss trials (Figure 3E). This was confirmed by a 2×4 repeated measures ANOVA with factors
CONDI-TION (hits, misses) and TIME showing main effects of condition [rPOC: F(1,20) = 69.94, p < 0.001; MFC: F(1,20) = 22.78, p < 0.001], but no significant inter-actions (rPOC: p = 0.177, MFC: p = 0.185).
Networks of cognitive control: Effects of time-on-task on inter-site phase synchronization
Our second hypothesis stated that top-down control of attention depends on communication between brain areas involved in attention, stimulus percep-tion, and cognitive control. We predicted that time-on-task related impairments would be manifested as diminished functional connectivity, as evidenced by decreases in inter-site phase synchronization. Be-cause of the prior theta phase coherence effects found in MFC, which is often regarded as the seat of cognitive control (Ridderinkhof et al., 2004), and in rPOC, which is known to be involved in spatial atten-tion and visual percepatten-tion (Thut et al., 2006), we first investigated long-rage phase synchronization in the theta band between rPOC and MFC.
CR trials averaged across all blocks showed a strong increase in phase synchronization confined to the same time-frequency window as the theta phase coherence effect (Figure 4A). Although there ap-peared to be substantial differences in phase syn-chronization magnitude between blocks
(
Figure 4B), this effect failed to reach significance [main effect of TIME, F(3.28,56.66) = 1.75, p = 0.161]. Despite the absence of a main effect of time, the development of the phase synchronization response over the course of the task again bore a distinct resemblance to the pattern of changes in behavioral performance (Figure 4C, cf. Figure 2A). We therefore computed correlations between theta-band phase synchroniza-tion and 𝐴′, but found no significant relation [𝑟̅ =0.11, p = 0.048 n.s. (see Statistical analyses)]. In addition to post-stimulus changes in phase synchronization, we were also interested in anticipa-tory effects that might drive changes in stimulus processing. There appeared to be traces of prepara-tory activity in the theta band (Figure 4A) starting at
about -600 ms before stimulus onset, but any differ-ences between blocks herein (Figure 4B) were not statistically significant [main effect of TIME, F(7,140) = 0.53, p = 0.814]. Moreover, we also performed an exploratory whole-head analysis of pre-stimulus alpha band-phase synchronization seeded from the rPOC. In this analysis, repeated measures ANOVAs were computed for every channel, with a Bonferroni correction for multiple comparisons as large as the number of tests. In particular, we presumed that other areas part of the frontoparietal attention net-work such as the intraparietal sulcus and frontal eye fields (Buschman & Miller, 2007; Capotosto et al.,
2009) may have also been recruited during the task,
but no effects survived the multiple comparison cor-rection.
Spatial attention orienting: Effects of time-on-task on pre-stimulus alpha dynamics
Third, we hypothesized that the time-on-task effect negatively impacts the capacity to spatially focus attention. The electrophysiological hallmark of ori-enting spatial attention is suppression of parieto-occipital alpha-band power over the processing hemisphere (in our case, rPOC), accompanied by an increase in alpha power over equivalent areas in the non-processing hemisphere (i.e. left POC) (Sauseng
et al., 2005; Thut et al., 2006). We therefore
predict-ed that this alpha power asymmetry would disappear following prolonged performance on the sustained attention task.
The condition-average time-frequency map of rPOC indeed shows an anticipatory effect between 8-12 Hz and from approximately -600 ms up until the time of stimulus presentation (Figure 5A1). Notably, while alpha power was suppressed in this baseline window during the first 10 minutes of the task, posi-tive values of alpha power were observed for the remainder of the task (Figure 5B1). However, the same plot also demonstrates that this difference was not confined to the time-window of interest but tonic in nature. Because of the absence of the expected
Figure 3 | Intertrial-phase coherence from 150 to 450 ms in the theta (4-7 Hz) frequency band changes as a function of time on task and is linked to behavioral performance. A through E show data from correct rejection trials. (A) Condition-average topograph-ical distribution of post-stimulus theta phase coherence. Electrodes PO8, P6 and P8 (gray circles) were pooled together for further analyses as the right parieto-occipital cortex (rPOC); the mid-frontal cortex is constituted by FCz, F1 and F2 (white circles). (B) Power within the same time-frequency window. Note the resemblance to phase coherence. (C) Condition average time-frequency maps of theta phase coherence hotspots (C1: rPOC, C2: MFC) with the time-frequency window of interest superimposed. (D) Line plots of theta phase coherence in separate time bins of correct rejection trials show the magnitude of the response changes with time-on-task, in both rPOC (D1) and MFC (D2). “CR10” = correct rejection trials in block 1, etc. Dashed lines show time range of window of interest. (E) Peak theta phase coherence within the window of interest plotted for each block. Dashed line marks the time of the motivation manipulation. Note that the time-course of peak phase coherence in both rPOC and MFC is distinctly similar to changes in behavioral performance shown in Figure 2A. (F) Peak theta phase coherence at both sites is significantly higher in hit than in miss trials.
rPOC (4−7 HZ)
Time (ms)
Phase coherence (a.u.)
−1000 −500 0 500 1000 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 CR10 CR20 CR30 CR40 CR50 CR60 CR70 CR 80 MFC (4−7 Hz) Time (ms)
Phase coherence (a.u.)
−1000 −500 0 500 1000 0.05 0.1 0.15 0.2 0.25 0.3 0.35 CR10 CR20 CR30 CR40 CR50 CR60 CR70 CR80 4−7 Hz; 150−450 ms Power (dB) −1 0 1 4−7 Hz; 150−450 ms
Phase coherence (a.u.)
0 0.2 0.4 1 2 3 4 5 6 7 8 0.3 0.35 0.4 0.45 4−7 Hz; 150−450 ms Block (10 minutes)
Phase coherence (a.u.)
rPOC MFC Time (ms) Frequency (Hz) rPOC −10002 −500 0 500 1000 3 6 10 16 28 47 80
Phase coherence (a.u.)
0 0.2 0.4 MFC Time (ms) Frequency (Hz) −10002 −500 0 500 1000 3 6 10 16 28 47 80
Phase coherence (a.u.)
0 0.2 0.4 rPOC MFC 0 0.1 0.2 0.3 0.4 0.5 0.6
Phase coherence (a.u.
) 4−7 Hz; 150−450 ms Hits Misses
B
D1
D2
C2
C1
A
E
F
phasic alpha suppression response, we opted in-stead to examine tonic changes in alpha power be-tween blocks over the entire course of the epoch (i.e. in a -1000 ms to +1000 ms window, with a con-dition-average baseline taken from the same time window) (Figure 5C1). Statistical tests on data within this new time window captured a marginally signifi-cant effect of time-on-task on alpha power [main effect of TIME: F(2.92,51.42) = 2.54, p = 0.067]. To
evaluate whether this alpha response was spatially selective, we identified the three equivalent elec-trodes to PO8, P6 and P8 (that together make up the rPOC pool) in the left hemisphere, which are PO7, P7 and P5 (henceforth: lPOC). However, again in contrast to our hypothesis, a 2×4 repeated measures ANOVA with factors HEMISPHERE (lPOC, rPOC) and TIME yielded no significant differ-ences [hemisphere: F(1,20) = 1.51, p = 0.233; time:
| Inter-site phase synchronization on correct rejection trials in the theta band (4-7 Hz) between mid-frontal and right
Figure 4
parieto-occipital cortices. (A) Condition average time-frequency map of phase synchronization, expressed as a % change from base-line (-1000 to -800 ms). A strong stimulus-evoked burst of synchrony occurred within the same time-frequency window (dashed base-line) as the theta phase coherence and power effects (see Figure 3). In addition, an increase in preparatory activity limited to the theta frequen-cy band is visible (dash-dots). (B) Line plots of theta-band phase synchronization in separate time bins. “CR10” = correct rejection trials in block 1, etc.; Dashed lines show time windows of interest. In spite of the clearly visible differences between blocks, no effects of time on phase synchronization reached significance (-600-0 ms: p = 0.814; 150-450 ms: p = 0.161). (C) Progression of peak post-stimulus theta phase synchronization over the course of the experiment. Though the shape of the curve bears a clear resemblance to both ′
(see Figure 2A) and thereby theta phase coherence (E), no significant correlations were obtained.
C
rPOC MFC Time (ms) Frequency (Hz) −10002 −500 0 500 1000 3 6 10 16 28 47 80 Synchrony change (%) −50 0 50B
A
Time (ms) Synchrony change (%) rPOC MFC (4−7 Hz) −1000 −500 0 500 1000 −10 0 10 20 30 40 50 60 70 80 90 CR10 CR20 CR30 CR40 CR50 CR60 CR70 CR80 1 2 3 4 5 6 7 8 50 60 70 80 90 100 110 120 130 rPOC MFC (4−7 Hz; 150−450 ms) Block (10 minutes) Synchrony change (%)F(2.96, 59.17) = 2.27, p = 0.090, interaction: F(7,140) = 0.95, p = 0.469]. Lastly, to identify possi-ble relations between tonic alpha power and behav-ioral performance, we (1) correlated alpha power changes between block 1 and 6 to concurrent changes in 𝐴′ and (2) theta phase coherence, and
(3) tested for differences between hit and miss tri-als.. Of these three, only the latter proved to be sig-nificant (main effect of CONDITION, F(1,20) = 14.69, p = 0.001) (Figure 5D).
A also shows a clear suppression of alpha Figure 5
oscillations later in time, which is typically observed post-stimulus and has previously been related to task effort (Palva & Palva, 2011). We therefore also examined whether this effect was subject to time-on-task related changes. However, analyses of 8-14 Hz power from 200-600 ms yielded no significant results (note: the window only envelops the lower part of the contour because the scalp topography of the higher frequency activity was not specific to rPOC).
Functional significance of slow oscillations in the time-on-task effect
In addition to the time-frequency results described above, we performed a separate convolution geared to expose effects of low-frequency (0.1-4 Hz) EEG fluctuations. For CR trials, we noted the presence of a prominent negativity over the rPOC in the delta frequency range between 1-3 Hz and -500 to 0 ms pre-stimulus (Figure 5A2). Upon closer examination, differences between conditions in this delta power response persisted throughout the trial, much like the tonic alpha power effect observed earlier
(
FigureB2, cf.
5 Figure 5B1). Therefore, we again used a time window enveloping the entire epoch for statisti-cal analyses, which revealed a significant effect of time-on-task on tonic delta power [main effect of TIME, F(2.09,41.71) = 6.56, p = 0.003]. The phasic delta power response did not appear to change over time, but the polarity of the effect did: delta power in the first few blocks was positive compared to base-line, but progressively decreased with time (Figure
C2). The motivation manipulation did not appear to 5
have a clear impact on delta power. As with alpha power, we correlated this delta suppression effect with 𝐴′ and theta phase coherence, and examined
hit and miss trials. For the latter, we also found a significant main effect of time [F(1.72, 34.44) = 3.91, p = 0.035], but no main effect of condition (p = 0.203) or a significant interaction (p = 0.371). Delta power was not associated with behavioral perfor-mance, but we did obtain a significant correlation between the magnitude of the delta suppression effect and theta phase coherence decline (r = -0.55, p = 0.0106).That is, those participants that exhibited the largest delta suppression after 60 minutes com-pared to the first 10 minutes, also had the smallest decline (or even a modest increase) in theta phase coherence (Figure 5E). This seems to suggest that the presence of delta oscillations may modulate the attentional system in such a way that the trial-to-trial variability of responses is affected. Because the motivation manipulation did not appear to influence tonic delta power, we used regular Spearman corre-lations instead of the correlation scheme optimized for repeated measures (see Statistical analyses).
Discussion
The present study aimed to further elucidate the role of top-down control and the neural correlates of at-tentional resources in the time-on-task effect. Our behavioral results showed that perceptual sensitivity declined as a function of time on task, and was only briefly susceptible to an increase in motivation. Se-cond, we found that theta phase coherence over right parieto-occipital (rPOC) and midfrontal cortex (MFC) was a marker of attentional stability and closely followed both the performance decrement and the response to the motivation manipulation. Third, there were some indications that the time-on-task effect is associated with changes in functional connectivity between these areas, as indexed by
| Tonic alpha (8-12 Hz) and delta (1-3 Hz) -band power change as a function of time-on-task and are linked to
behav-Figure 5
ioral performance and attentional stability. A through C and E show data from correct rejection trials. (A) Condition-average time-frequency plots of high- (A1) and low-time-frequency (A2) power at right parieto-occipital electrodes, with windows of interest superimposed [for theta power results (dashed line), see text; no significant results were obtained for 8-14 Hz power from 200 to 600 ms (dash-dots)]. (B) Line plots of power in separate time bins, with a condition-average baseline from -1000 to -800 ms (“CR10” = correct rejection trials in block 1, etc.). Suppression of alpha power (B1) virtually disappears after merely 10 minutes of sustained attention; whereas suppres-sion of delta power (B2) increases progressively with time-on-task. (C) Tonic changes in alpha (C1) and delta power (C2) over the course of the experiment per 10-minute block. Because of the tonic nature of power changes (see B), the baseline was chosen to en-velop the entire trial. Dashed lines show the time of the motivation instruction; asterisks demonstrate significance at p < 0.05 compared to the first block. (D) Whole-epoch alpha power is significantly decreased in hit compared to miss trials (shared -1000-1000 ms base-line) (see text for other analyses). (E) Scatter plot of tonic delta power changes over time (1st block (10 min) minus 6th block (60 min);
x-axis) vs. peak phase coherence changes over the same time interval (y-x-axis). Each dot represents a single subject; Spearman’s rank-order correlation coefficient is shown (p = 0.0106) (see text for details and other analyses).
−2 −1 0 1 2 3 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2 0.25 0.3
Delta power change (dB)
Theta ITPC change (a.u.)
r = −0.55 rPOC Hits Misses −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 Power (dB) rPOC (8−12 Hz; −1000−1000 ms) 1 2 3 4 5 6 7 8 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 rPOC (1−3 Hz; −1000−1000 ms) Block (10 minutes) Power (dB) 1 2 3 4 5 6 7 8 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 rPOC (8−12 Hz; −1000−1000 ms) Block (10 minutes) Power (dB) rPOC (1−3 Hz) Time (ms) Power (dB) −1000−1 −500 0 500 1000 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 CR10 CR20 CR30 CR40 CR50 CR60 CR70 CR80 rPOC (8−12 Hz) Time (ms) Power (dB) −1000 −500 0 500 1000 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 rPOC Time (ms) Frequency (Hz) −1000 −500 0 500 1000 0.1 0.5 1 2 3 4 5 Power (dB) −1 0 1 Time (ms) Frequency (Hz) rPOC −10002 −500 0 500 1000 3 6 10 16 28 47 80
A2
C1
C2
B2
B1
A1
D
E
theta-band phase synchronization. Finally, our re-sults displayed tonic changes in alpha- and delta power that were linked to target perception and at-tentional stability, respectively. We will now discuss our interpretation of these findings in more detail. Effects of time-on-task and motivation on perceptual sensitivity
We first hypothesized that perceptual sensitivity, expressed as 𝐴′ scores, would decrease with
time-on-task. Indeed, we found that perceptual sensitivity was subject to the classic performance decrement (e.g. Parasuraman (1979)).
In addition, we confirmed our prediction that once performance has gone down, increases in motivation should not be able to fully reverse this trend (Helton & Warm, 2008). Subjective ratings indicated that the monetary incentive successfully heightened motivation to perform well and lowered aversion towards the task. The motivation manipula-tion may have also exerted a transient positive effect on performance, since perceptual sensitivity in the 10 minutes thereafter was elevated visibly. Never-theless, the most compelling conclusion is that moti-vation was on the whole ineffective at rescuing per-formance, for the following three reasons. First, de-spite the apparent boost, 𝐴′ scores after the
motiva-tion manipulamotiva-tion were still significantly lower than those in the first block. Second, in the final ten minutes of the experiment, perceptual sensitivity dropped to the lowest score overall, even though participants were clearly instructed that in order to receive the extra sum of cash, they had to maintain good performance rates for the full twenty minutes. Finally, the motivation ratings after the monetary incentive did not significantly differ from the first block, indicating that subjects were still as motivated to perform well, yet they were unable to keep up performance in the last block of the task.
All in all, we conclude that our present results are congruent with resource theory’s proposal that the performance decrement stems from depletion of
cognitive resources (Helton & Warm, 2008). This conclusion is relevant for the interpretation of our EEG results, but since our aim was not to explicitly test different theories of the time-on-task effect, we cannot be confident that mindlessness does not contribute to the performance decrement at all. In-terestingly, Smallwood (2010) notes that theories of mindlessness and cognitive resources are not mutu-ally exclusive. After all, mind wandering is also re-source demanding in the sense that it is a con-sciously reportable experience that occupies part of the global workspace. In the case of vigilance, an increase in task-unrelated thoughts due to the time-on-task effect may thus accelerate depletion of cog-nitive resources.
Stimulus-locked theta oscillations un-derlie target perception
Based on our previous work (Lutz et al., 2009), we hypothesized that prolonged time-on-task would be associated with increased variability in stimulus-induced responses as reflected by declining theta phase coherence, and that such changes would thus be associated with the behavioral performance dec-rement. This previous study implicated theta phase coherence in conscious target perception and sus-tained attention, through the finding that meditation practice improved performance, reduced reaction time variability and strengthened theta-band phase coherence over anterior areas. These results demonstrate the plasticity of both cognitive process-es such as sustained attention, and the underlying brain responses. Here we show the occurrence of plasticity in the opposite direction: whereas mental training improves attentional stability over time, men-tal fatigue with time on task decreases the trial-to-trial stability of stimulus processing. The combination of these results illustrates that cognitive and neural processes are not stable entities but highly suscep-tible to change, both for better and for worse.
One notable difference between the present experiment and the mental training study is that
while the latter found effects of theta phase coher-ence over anterior areas only, our results showed significant activity over both mid-frontal and right parieto-occipital electrodes. At both sites, changes in phase coherence very closely shadowed the trend in behavioral performance, which stands out from eye-balling the graphs alone (cf. Figures 2A and 3E). This tight coupling between 𝐴′ and theta phase
co-herence suggests that attentional stability plays a major role in the time-on-task effect. In addition, theta phase coherence can be seen as a suitable candidate for an electrophysiological correlate of top-down cognitive resources that get depleted dur-ing sustained attention.
Theta-band phase coherence from 150 to 450 ms post-stimulus was accompanied by significant increases in theta power in the same time window. The presence of concurrent changes in theta power suggests that the theta phase coherence effect was due to stimulus-evoked neural activity and not phase resetting of ongoing oscillations (Sauseng &
Klimesch, 2008).
The theta rhythm in general has long been known to increase following stimulus presentation (e.g. (Demiralp & Başar, 1992)). Theta phase coher-ence in particular has received much less attention, though it was recently proposed that theta-phase locking can modulate the visual P2 ERP component
(Freunberger et al., 2007). That same study showed
that parieto-occipital generators give rise to the P2. Moreover, both the P2 and theta phase have been associated with a memory matching process that compares sensory input to stored target representa-tions (Rizzuto et al., 2006). Though we did not pres-ently perform any analyses geared to event-related potentials, this interpretation fits well with our task and the general importance of theta oscillations in memory.
Cognitive control and long-range neu-ral communication
Time-on-task related changes in neural activity are most likely not restricted to local dynamics, but should be manifested on the network level as well. We reported that inter-site phase synchronization between right parieto-occipital and mid-frontal cortex appears to change over time, and that the pattern of these changes strongly resembles that of behavioral performance. While certainly interesting, both results fell short of reaching statistical significance. We also performed a whole-head phase synchronization analysis seeded from rPOC, but this also did not yield any significant results. At present it is therefore difficult to meaningfully interpret these results, but they are worth exploring further. It seems quite plau-sible that MFC and rPOC would stand in communi-cation with each other, considering their functional importance in cognitive control and visuospatial at-tention, respectively (Ridderinkhof et al., 2004; Thut et al., 2006).
Though inter-site phase synchronization does not give any information about directionality, their respective roles in higher-order cognition and per-ception would credit an interpretation that MFC is driving the rPOC. In this regard it is interesting that mid-frontal theta phase coherence appears to dis-play a relatively larger increase after the motivation manipulation than rPOC. This could be interpreted as a signal originating in the MFC, by which the rPOC is “instructed” to escalate efforts towards iden-tification of visual stimuli. However, MFC-rPOC communication could also occur in the reverse direc-tion. A recent study employed Granger causality to demonstrate that during anticipation of visual stimuli, alpha-band functional connectivity was directed from occipital to mid-frontal areas (Cohen & van Gaal, 2012).
Anticipatory alpha rhythms and top-down control of spatial attention
Pre-stimulus alpha-band power over parieto-occipital areas is known to be sensitive to upcoming visual events. Following a covert shift in spatial attention to the left or right hemifield, posterior alpha power is enhanced in the (non-processing) ipsilateral hemi-sphere and suppressed in the (processing) contrala-teral hemisphere (Sauseng et al., 2005; Thut et al.,
2006). We set out to replicate this effect and to
demonstrate that the hemispheric alpha asymmetry would decline with time-on-task, possibly reflecting depletion of resources of top-down control. Our re-sults show a marked decrease in anticipatory alpha suppression already after 10 minutes of sustained attention, even changing polarity to an increase in alpha power. However, this was in fact not a phasic but a tonic response that persisted throughout the trial. In addition, we found no significant differences in modulation of alpha power between hemispheres.
Thereby, the present experiment thus failed to replicate the spatial biasing of endogenous alpha oscillations. We distinguish two main reasons that may explain the absence of this effect in our data. First, studies that previously reported this effect (e.g.
Sauseng et al. (2005); Thut et al. (2006)) employed
paradigms that differ from the sustained attention task we used. Generally, these tasks invoked clear visual or auditory cues that directed attention to ei-ther hemifield, followed by presentation of the stimu-lus after a delay period. In contrast, our task re-quired participants to constantly focus their attention on the same spot in the same hemifield for a pro-longed period of time. It is plausible that the spatial attention bias only manifests itself during these short and cued shifts in spatial attention, but not during sustained attention. This interpretation supports a conception of attention as a multi-faceted phenome-non, with sustained and short-term attention being fundamentally different. A second, more methodo-logical reason could lie within the difficulties
associ-ated with choosing an appropriate baseline for the EEG data. Often, researchers are interested in post-stimulus activity and can safely choose a pre-stimulus baseline. However, since our interval of interest for alpha power dynamics was also prior to stimulus presentation, it is possible that the baselin-ing procedure partly obscured these anticipatory effects.
Low-frequency dynamics modulate post-stimulus theta phase coherence Apart from tonic alpha power, we also found a signif-icant effect of time-on-task on oscillatory power in the delta (1-3 Hz) band: as time-on-task increased, tonic delta power grew increasingly more negative. This change in delta power magnitude from 10 minutes to 60 minutes after the start of the experi-ment was correlated to concurrent changes in post-stimulus theta phase coherence. More specifically, those subjects who exhibited the smallest reduction in theta phase coherence after 60 minutes also dis-played the largest suppression of delta power. Since theta phase coherence proved to be a close corre-late of behavioral performance, this result implies that tonic reductions in delta power are beneficial. However, since delta suppression increased with time-on-task, another interpretation would be that it is detrimental. More generally, the finding that delta-band activity decreased with time-on-task is in itself surprising, since slow oscillations play a prominent role in sleep (Steriade, 2006).
These findings may seem paradoxical at first, but form a more coherent picture when considering the complexities of the sustained attention paradigm. Successful task performance can be due to use of exogenous cues (i.e. the mask appearing right be-fore the target stimulus), but is more efficiently at-tained by using the endogenous cue that targets are always spaced exactly 2000 ms apart. These strate-gies roughly map onto the distinction between reac-tive and proacreac-tive control, respecreac-tively (Braver et al.,
least, the brain operates in a rhythmic mode of pro-cessing wherein the-phase locking of neuronal re-sponses is made possible by entrainment to slow oscillations (Jacobs et al., 2007). However, this strategy requires top-down control of attention, which may suffer from resource depletion over time. Correspondingly, Schroeder and Lakatos (2009)
suggest that during vigilance, the brain engages in a different mode of operation involving suppression of low-frequency oscillations.
Limitations & future directions
Our experiment revealed a complex system of changes in pre-stimulus and post-stimulus activity, in theta, alpha and delta frequency bands, and in the dimensions of power, phase coherence and phase synchronization. Future studies could focus on ex-plicitly testing a subset of the conclusions drawn here. For instance, if delta oscillations are indeed suppressed by a transition from top-down control into a continuous state of excitability, one would expect that a brief break in attentional focus would lead to replenishment of resources accompanied by increases in delta-band power (Schroeder &
Lakatos, 2009). A second limitation of the present
study is that due to insufficient hit and miss trials, only CR trials could be analyzed in 10 minute bins, and analysis of FA trials was not even possible. FA trials might prove particularly useful to evaluate the role of pre-stimulus alpha oscillations, since the oc-currence of a false alarm likely stems from mishaps in anticipatory orienting of attention (Mazaheri et al.,
2009). Furthermore, our current design lacked a
means with which to assess the relative contribution of mindlessness to the time-on-task effect. It would be informative to have subjects report the extent of task-unrelated-thoughts during the experiment, much like the motivation/aversion ratings. Finally, this complex design also warrants the use of more sophisticated statistical analyses that accommodate the repeated measures structure of the data. Instead of performing individual correlations between two
measures that both change over time, developing a mixed-modeling approach that captures both within- and between-subject variance such as suggested by
Roy (2006) might be more appropriate.
Conclusions
The present study provides an in-depth investigation of the neural and psychological correlates of mental fatigue during sustained attention. We found that the time-on-task effect: (1) produced a performance decrement that cannot be solely attributed to loss of motivation, (2) increased variability in attentional processing of visual stimuli, (3) possibly reduced functional connectivity between cognitive control and visuospatial attention-related brain areas, and (4) induced tonic changes in local power that in turn modulated target perception and stimulus-locked responses. These effects are currently best ex-plained in terms of a depletion of resources of top-down control, which causes a shift to a different mode of attentional processing in the brain. With further research, these findings may pave the way for practical applications, such as early detection of fatigue-related performance decrements in people operating in dangerous environments, or combatting fatigue through transcranial electrical stimulation techniques.
Acknowledgements
I would primarily like to thank Heleen Slagter for making this study possible in the first place, for her patience and for invaluable comments from beginning to end. Data collection was all done by Sam Prinssen with assistance from Rudy van den Brink; I am especially grateful to the latter for helping me familiarize myself with the data. Last-ly, I would like to express my gratitude to Mike Cohen, who wrote most of the code and taught me nearly all I know about EEG data analysis.
