Conference PaperPDF Available

Gamma Power as an Index of Sustained Attention in Simulated Vigilance Tasks

Authors:

Abstract and Figures

Performance on the psychomotor vigilance test (PVT; Dinges and Powell, 1985) - a common index of sustained attention - is affected by the opposing forces of fatigue and sustained effort, where reaction times and error rates typically increase across trials and are sometimes offset by additional efforts deployed toward the end of the task (i.e., an "end-spurt"; c.f. Bergum and Klein, 1961). In ACT-R (Adaptive Control of Thought-Rational; Anderson et al., 2004), these influences on task performance have been modeled as latent variables that are inferred from performance (e.g., Jongman, 1998; Vek-sler and Gunzelmann, 2018) without connections to directly observable variables. We propose the use of frontal gamma (γ) spectral power as a direct measure of vigilant effort and demonstrate its efficacy in modeling performance on the PVT in both the aggregate and in individuals.
Content may be subject to copyright.
Gamma Power as an Index of Sustained Attention in Simulated Vigilance Tasks
Taylor Curley (taylor.curley@cubic.com)
Cubic Defense
Beavercreek, OH 45324 USA
Lorraine Borghetti (lorraine.borghetti.1.ctr@us.af.mil)
Air Force Research Laboratory
WPAFB, OH 45433 USA
Megan B. Morris (megan.morris.3@us.af.mil)
Air Force Research Laboratory
WPAFB, OH 45433 USA
Abstract
Performance on the psychomotor vigilance test (PVT; Dinges
and Powell, 1985)—a common index of sustained atten-
tion—is affected by the opposing forces of fatigue and sus-
tained effort, where reaction times and error rates typically in-
crease across trials and are sometimes offset by additional ef-
forts deployed toward the end of the task (i.e., an “end-spurt”;
c.f. Bergum and Klein, 1961). In ACT-R (Adaptive Control
of Thought-Rational; Anderson et al., 2004), these influences
on task performance have been modeled as latent variables
that are inferred from performance (e.g., Jongman, 1998; Vek-
sler and Gunzelmann, 2018) without connections to directly
observable variables. We propose the use of frontal gamma
(γ) spectral power as a direct measure of vigilant effort and
demonstrate its efficacy in modeling performance on the PVT
in both the aggregate and in individuals.
Keywords: ACT-R; EEG; fatigue; vigilance; microlapse
Introduction
A well-documented phenomenon in human performance re-
search is the decline in performance during extended vigi-
lance tasks due to cognitive and physical fatigue (c.f., Ack-
erman, 2011). The relative simplicity of common sustained
attention tasks, such as the psychomotor vigilance test (PVT;
Dinges and Powell, 1985), however, overshadows the com-
plex and arcane connections between task outcomes and the
neural mechanisms that give rise to these outcomes (Ishii
et al., 2014; Kim et al., 2017). Despite this, changes in
electroencephalographic (EEG) activity have been shown to
provide a potentially reliable marker of mental fatigue (Tran
et al., 2020).
One way to examine links between cognitive and neural
mechanisms of sustained attention is by integrating data from
behavioral and neural sources into a single model (Turner
et al., 2017). In the ACT-R (Adaptive Control of Thought-
Rational; Anderson et al., 2004) cognitive architecture, for
example, researchers have begun to use event-related poten-
tials (ERPs; Cassenti et al., 2011) and neural “blips” (Borst
and Anderson, 2015) to link selection and duration of indi-
vidual behaviors (productions) to EEG data. Despite exten-
sive work on modeling the effects of time-on-task (Veksler
and Gunzelmann, 2018) and sleep deprivation (Gunzelmann
et al., 2009, 2015) on the PVT, ACT-R practitioners have yet
2.0
2.1
2.2
2.3
2.4
2.5
12345
Time Bin
Gamma Power
Source
Fz
Pz
Figure 1: Observed relative gamma spectral power density
across 2 minute time bins during the PVT. From Borghetti
et al. (2021)1.
to directly investigate the use of EEG in modeling fatigue-
related decrements during vigilance tasks.
We propose the use of estimated power in frontal gamma
(γ) wave forms in models of vigilant attention. Specifically,
we argue that γpower measured during the PVT is a reliable
index of sustained attention that reflects fatigue (e.g., perfor-
mance decreases across time) and compensation (e.g., end-
spurts) and can be directly applied to ACT-R parameters. To
this end, we first review relevant investigations of EEG and
ACT-R as they relate to vigilance and then introduce a method
for incorporating γpower into ACT-R models of the PVT.
EEG and Fatigue
Recently, Borghetti et al. (2021) reported a study examining
electrophysiological measurements from 34 young adult par-
ticipants (Mage =22.6) over the course of a 10-min PVT in
which participants were asked to respond immediately when
a stimulus appears on the screen. Vigilance decrements dur-
ing the PVT were exemplified by positive shifts in the dis-
tributions of reaction times, indicating increasingly slower
responses, as well as increases in premature responses, i.e.,
false alarms (Doran et al., 2001). The results of the behavioral
task also show a slight improvement in task performance in
later trials, indicating an increase in effort, i.e. an “end-spurt”
(e.g., Bergum and Klein, 1961).
The authors examined spectral power density, or an esti-
mate of the power in a neural signal given a particular fre-
quency, over the course of the 10-min task, focusing on theta
(θ, 3-8 Hz), alpha (α, 9-14 Hz), beta (β, 15-30 Hz), and
gamma (γ, 30-100 Hz) wave forms1. The top half of Fig-
ure 1 illustrates the main findings of the study: Significant
trends indicating decreases in γspectral power across time-
on-task in both the frontal (Fz) and parietal (Pz) regions of
the brain, with a significant end-spurt towards the end of the
task (Morris et al., 2020). Borghetti et al. (2021) concluded
that frontal γindexes the dynamic between fatigue and sus-
tained attention in the PVT. This is consistent with similar re-
search indicating increases in γactivity across vigilance tasks
(Kim et al., 2017) and positive associations between task per-
formance and amplitudes of γoscillations (Herrmann et al.,
2010).
Fatigue and Compensatory Effort in ACT-R
The ACT-R cognitive architecture provides a rich environ-
ment for investigating effort and fatigue in goal-driven tasks,
where influences on effort during the task are modeled as pa-
rameters affecting the selection and execution of procedural
knowledge, i.e., “productions”. During the course of the task,
the model selects productions with the greatest estimated util-
ities (U), or a parameter indicating the strength and appropri-
ateness of a given behavior at a given time. In prior versions
of ACT-R, utilities were determined by the probability that a
given goal will lead to success (P), the value of the current
goal (G), and the cost of using that particular production to
reach a goal (C). In the current version, production selection
is a function of an initial utility value parameter (υ), noise
on this value (σ2), and a threshold parameter (τ), wherein the
model selects the production with the highest above-threshold
utility value to fire. Production utility values can either re-
main static or can update to reflect changes in the model’s en-
vironment, such as production learning/reinforcement (e.g.,
Lovett and Anderson, 1996).
Previous studies have conceptualized vigilant effort as a di-
rect influence on production utilities. Jongman (1998), for ex-
ample, used parameterized “motivation” in a previous ACT-
R architecture to directly influence G, where greater Gval-
ues represent greater effort allocated toward achieving a goal
and lead to better task outcomes, but lower Gvalues result in
firing inappropriate productions. Belavkin (2001) also used
Gto influence utility values, but conceptualized the param-
eter as reflecting a more general “arousal” state, where de-
creases in Gresult in fewer above-threshold productions, re-
sulting in “giving-up” behavior. In contrast, Gunzelmann
et al. (2009) simulated fatigue by imparting its effects on both
utility values (through the Gparameter) and τas a function of
“arousal” (A), which is derived from biomathematical esti-
1These results are based on a correction to the gamma spectral
power analyses. In the original version of the paper, gamma es-
timates decreased sharply between 0 and 2 minutes and declined
slightly across minutes 2 and 10. The corrected analyses indicate
that gamma power increases between time bins 4 (6 - 8 m) and 5 (8
- 10 m), as shown in Figure 1.
mates of arousal (c.f. Van Dongen, 2004). The decrease in
utility and τvalues represent the deleterious effects of fatigue
and efforts enacted to compensate for fatigue, respectively.
Gunzelmann et al. (2009) also incorporated “microlapses”,
or simulated lapses in attention. Microlapses occur when the
utility module is unable to select a production, such as when
all utility values are lower than τ. The occurrence of a micro-
lapse results in a penalty to utility values and thus increases
the probability of future microlapses. While the number of
microlapses that occur during a simulated task is not con-
trolled by the modeler, the penalty to utility values can be
freely-estimated.
More recently, Veksler and Gunzelmann (2018) general-
ized decrements in arousal as stemming from the effects of
time spent engaging in the task (“time-on-task”) and simu-
lated microlapses. Specifically, the authors estimate the util-
ity of a production Uby imposing a penalty on the initial
production utility value (υ) as a function of both the number
of microlapses (Nml) and the time spent on the experiment (t):
U(t) = υ[λNml (1+t)ρ],(1)
where υis the initial utility value parameter, tis time spent
on the task (scaled to minutes), λscales the effect of micro-
lapses on utility values, and ρscales the effect of time-on-
task. As fatigue increases and production values decrease,
the probability of sampling an inappropriate production in-
creases, leading to increases in false alarms.
In contrast, the production utility selection threshold is
only affected by time-on-task:
U T (t) = τ(1+t)κ,(2)
where τis the initial utility theshold parameter and κscales
the effect of time-on-task on the threshold. Lower thresh-
olds under conditions of fatigue allow the model to select pro-
ductions whose υvalues have decreased. This compensation
is imperfect, however, as lowering the production selection
threshold also allows the model to fire productions that are
not appropriate for the context. In models of the PVT, this
leads to increases in false starts and misses.
Candidates for Integration
We now review mechanisms for 1) translating γspectral
power to units appropriate for use in ACT-R simulations and
2) applying transformed γestimates to the ACT-R cognitive
architecture.
Scaling Spectral Power Estimates. Similar to previous re-
search (e.g., Belavkin, 2001; Gunzelmann et al., 2009; Jong-
man, 1998; Veksler and Gunzelmann, 2018), we conceptu-
alize sustained attention as a parameter ζthat is typically
bounded between zero and one. In the proposed model, how-
ever, ζcan occasionally exceed its upper bound, meaning that
parameterized effort cannot go below zero (meaning “abso-
lute” fatigue), but can surpass unity (meaning “extra” effort).
Param. Description Bounds Value ζModel?
υInitial production utility value [0.0,Inf ]Free Yes
τInitial production utility theshold [0.0,Inf ]Free Yes
ρProduction utility time-on-task penalty [1.0,0.0]Free No
κUtility threshold time-on-task penalty [1.0,0.0]Free No
λMicrolapse penalty [0.0,1.0]Free Yes
φConflict resolution time N/A 0.05 Yes
Table 1: Descriptions of fatigue-related parameters in the ACT-R model of the PVT. The “Value” column indicates if a value
is freely-estimated, and if not, what the value is fixed to. The ζModel?” column indicates if the parameter is included in the
model that uses γpower as a performance moderator. All 6 parameters are included in the full (“Fatigue”) model (c.f. Veksler
and Gunzelmann, 2018).
Thus, ζcan capture decrements due to fatigue as well as com-
pensatory efforts that offset fatigue, such as the end-spurt ef-
fect (Morris et al., 2020).
One way to normalize fatigue moderator values is by ad-
justing the values to the smallest value and the range of the
values. This normalization method has been used to scale
biomathematical estimates of arousal in previous investiga-
tions of the PVT (Gunzelmann et al., 2009), where estimates
start with high values and monotonically decrease as a func-
tion of time. An interesting aspect of this method that is re-
flected in the fatigue moderators proposed by Gunzelmann
and colleagues (Gunzelmann et al., 2009; Veksler and Gun-
zelmann, 2018) is that the normalized values start at 1 (the
highest possible value) and decrease with time-on-task, im-
plying that performance cannot meet or exceed that from
t=1. Therefore, we opted to normalize γito the first ob-
servation in order to simulate end-spurt effects.
Given a set of observed spectral power estimates (total or
relative) Γi={γi,1,...,γi,t}, for participant iat time t, as well
as the range of these values, γri=range{γi,1,...,γi,t}, we can
calculate effort as:
ζi,t=1+γi,tγi,1
γri.(3)
Here, ζi,1=1 and all subsequent values are interpreted as
diminished effort due to time-on-task (ζi,tζi,1) or additional
(i.e., compensatory) effort compared to baseline (ζi,tζi,1),
allowing the model to account for end-spurt effects.
Applying Fatigue Decrements. The theoretic interpreta-
tion of γwith respect to vigilance is intentionally vague
(i.e., an index of sustained attention) and does not allow for
a straightforward implementation of the ζparameter in the
ACT-R architecture. In these simulations, we integrate pa-
rameterized effort in a linear function with the initial produc-
tion utility parameter υ(similar to Eq 1) with brief lapses in
attention. Therefore, the modulated utility value at a given
time, U(t), can be calculated as a function of υ,ζ, and the
number of simulated microlapses (Nml):
U(t) = υ·[λNml ·ζi,t].(4)
The Current Study
The estimated penalties to utility values and thresholds in
ACT-R are imperfect. First, they are “smoothed” approxima-
tions of behavior and are unlikely to directly capture stochas-
tic, asynchronous intraindividual variability across time, lead-
ing to error inflation when fitting fatigue parameters to indi-
vidual participants. Second, these mechanisms are indirect
inferences resulting from observations of behavioral data and
have yet to be empirically linked to outside indicators.
The current project addresses these issues by examining the
extent to which neural indices of vigilance correspond to the
deleterious effects of fatigue in the PVT. Specifically, spectral
power density in γwaveforms is expected to accurately cap-
ture fatigue and effort in ACT-R models of task performance.
We expect to find that models using the observed power den-
sity estimates (Equations 3 and 4) in place of fatigue functions
(Equations 1 and 2) will fit the observed data as well as, if not
better than, models with these functions in both the aggregate
and at the level of the individual.
Methods
Thirty-four adult volunteers (Mage = 22.60; SDage = 4.08) re-
cruited through the University of Dayton Research Institute
(UDRI) participated in a single 2-h study session consisting
of 3 experiment tasks with simultaneous EEG recording. The
study was approved by institutional review boards at both
UDRI and the Air Force Research Laboratory (AFRL), and
all individuals were compensated for their participation in the
study.
We provide a quick overview of the behavioral and elec-
trophysiology methods below; further details can be found in
Borghetti et al. (2021).
Behavioral
Participants were asked to participate in a 10-m PVT task as
a part of the 2-h study session. During the PVT, participants
were asked to monitor a computer screen with a black back-
ground and to press “j” on a standard computer keyboard as
quickly as possible to a target stimulus, i.e., white numbers in
the middle of the screen displaying the time (in ms) since tar-
get onset. The time in between the previous response and the
0.315
0.320
0.325
0.330
0.335
12345
Time Bin
Mean RT (s)
0.03
0.04
0.05
0.06
0.07
0.08
1 2 3 4 5
Time Bin
p(Lapse)
Figure 2: Performance data by time bins for average RTs for
valid trials (left) and average proportion of lapses (right). Er-
ror bars represent the standard error of the mean.
onset of a new stimulus, the interstimulus interval (ISI), was
randomly selected from an interval between 2 and 10 s. ISIs
were exact integers and selected from a uniform distribution.
EEG
Briefly, participants were fitted with an EEG cap with 64
electrodes, with 2 flat, unlinked electrodes applied to the
mastoids. These data were processed using custom MAT-
LAB scripts along with the EEGLAB toolbox (Delorme and
Makeig, 2004). After applying a 1 Hz high-pass filter and
removing artifacts, these data were epoched into segments of
±1500 ms with respect to stimulus onset and divided into
five, 2-m time bins. For the gamma spectral analysis, we as-
sayed power in the 70-100 Hz frequency band for frontal (Fz)
and parietal (Pz) cortical regions.
Computational
The computational model was programmed using a Julia lan-
guage (Bezanson et al., 2017) implementation of the ACT-
R cognitive architecture (Anderson et al., 2004). In ACT-R,
the PVT has been modeled as a time-inhomogenous semi-
Markov process consisting of three phases (Gunzelmann
et al., 2009; Veksler and Gunzelmann, 2018): Wait,Attend,
and Respond. The Wait production occurs prior to stimulus
onset in anticipation of the next trial, while the Attend and
Respond productions occur after a critical stimulus has been
visually processed and after the decision has been made to en-
gage in a response, respectively. These productions typically
occur in the Wait-Attend-Respond sequence, but the order can
be disrupted if an inappropriate production is selected on the
basis of low utility values (U). This can lead to false starts,
where the Respond production is selected in the absence of
a valid stimulus (i.e., RTs <150 ms), and lapses, where the
model fails to select the Attend or Respond productions in
the presence of a valid stimulus (i.e., RTs >500 ms). Addi-
tionally, response latency is penalized whenever there are no
productions that exceed the production utility threshold (UT )
by adding 50 ms for each occurrence (microlapse; c.f. Gun-
zelmann et al., 2009).
Importantly, the ACT-R model of the PVT simulates fa-
tigue by applying a penalty to a) only initial utility values
(Jongman, 1998; Belavkin, 2001) or b) both initial utility val-
ues and utility thresholds (Gunzelmann et al., 2009, 2015;
1
2
3
4
5
12345
Est. Utility Value
Est. Utility Threshold
0
1
2
3
4
5
0.00 0.25 0.50 0.75 1.00
Est. Microlapse Penalty
Density
0e+00
5e05
1e04
0 2500 5000 7500 10000
AIC
Density
Figure 3: Best-fitting estimates for υand τ(left), λ(top right),
and associated AIC values (bottom right).
Veksler and Gunzelmann, 2018). Here, we only penalize util-
ity values derived from Equations 3 and 4 based on re-scaled
gamma power estimates. Table 1 provides descriptions of the
parameters, the ranges of possible values, and the models that
they are used in.
Results
Behavioral
We performed statistical analyses on responses categorized
into 3 types: False starts (RTs <150 ms), lapses (RTs >
500 ms), and valid responses (150 ms RTS 500 ms).
For computational ease, we binned the data into five, 2-m
bins and applied an inverse transformation to the RTs, i.e.,
1/(RT 1000)(Ratcliff, 1993).
A repeated measures ANOVA on the aggregated inverted
RT values with a Greenhouse-Geisser correction on the de-
grees of freedom (W= 0.53, p= 0.02) indicates that the ef-
fect of time bin is significant, F(2.89,98.41) = 11.54, p<
0.05, where average RTs increase between the first and fourth
time bins (i.e., minutes 0 - 8), but decrease slightly in the fifth
time bin (i.e., minutes 8 - 10; c.f. Figure 2). A similar one-
way logistic GLM on lapses indicates that the log-odds of this
type of response change across time bins, F(4,3652) = 3.48,
p<0.05, where lapse rates decrease between bins 1 and 2,
increase between bins 2 and 4, and then decrease again be-
tween bins 4 and 5 (c.f. Figure 2). A one-way logistic GLM
indicates that the probability of a false start on any given trial
is not different across time bins, F(4,3651) <0.1.
Spectral Power
For frontal γ(Figure 1), a Friedman test on total power
estimates across blocks is significant, χ2(4) = 11.3, p=
0.02. Follow-up paired comparisons indicate that estimates
increase significantly between Blocks 2 and 3, p<0.05, de-
crease significantly between Blocks 3 and 4, p<0.05, and
increase with marginal significance between blocks 4 and 5,
p= 0.07, although only the significance of the first compar-
ison survives after Bonferroni corrections to the degrees of
freedom.
Parameters Fit Indices
Estimate Model υ τ ρ κ λ -2LL AIC BIC
Aggregate Fatigue 4.01 2.90 -0.28 -0.20 0.98 5829.99 5839.99 5877.58
Gamma 3.15 2.07 - - 0.74 3920.76 3926.76 3949.25
Individual
Fatigue 5.78 0.32 -0.41 -0.17 0.81 6591.08 6601.08 6604.58
(0.51) (0.08) (0.05) (0.03) (0.03) (227.69) (227.69) (227.69)
Gamma 3.36 2.74 - - 0.73 4090.21 4096.21 4098.01
(0.08)(0.07) - - (0.01) (424.64) (424.64) (424.64)
Table 2: Best-fitting parameters for aggregated data (top) and summary statistics of the best-fitting parameters for individuals
(bottom). For individuals, we report the means and standard errors of the mean (in parentheses) of these estimates. “Fatigue”
refers to models using the decrement parameters described by Veksler and Gunzelmann (2018) while “Gamma” refers to the
proposed model.
Computational
We estimated the parameters for two different models—one
using the fatigue moderators described by Veksler and
Gunzelmann (2018) and another using gamma power esti-
mates—using the data from individual participants and ag-
gregated across all participants. Model fit was calculated us-
ing the summed log-likelihoods of the simulated RT data to
log-normal distributions based on the observed RTs. We used
a simplex search algorithm via Optim.jl (Mogensen and
Riseth, 2018) to find the parameter values that maximized the
likelihood of the two PVT models given the observed data.
We repeated the optimization procedure 15 times for each set
of data, using new starting values on each iteration to avoid
local minima. Table 2 details the best-fitting parameters by
data source (“Aggregate” vs. “Individual”) and by the type of
model (“Fatigue” vs. “Gamma”).
Overall, the model using gamma spectral power density as
a direct influence on utility values provides a better fit to the
observed data than the model using established computational
fatigue moderators. For the aggregated data, the difference
in fit statistics suggest that there is decisive evidence (Kass
and Raftery, 1995) in favor of the Gamma model, logB10 =
1928.33. Similarly, the difference in average fit values across
all participants for the two models also suggests that there
is decisive evidence in favor of the Gamma model, logB10 =
2506.57. Across individuals, the Gamma model is favored
over the Fatigue model for all but 5 of the 34 participants in
the study.
Discussion
In this paper, we introduced an ACT-R model of vigilant at-
tention that directly integrates frontal γspectral power density
estimates into the parameters of the model that influence task
performance. We compared the ability of the new model to
fit observed RT data to that of a similar model of PVT perfor-
mance and found that the proposed model provides a better fit
to both aggregated and individual data than previous models
of fatigue. These results suggest that frontal γpower esti-
mates can be used as a measure of sustained attention and
Figure 4: Reaction time distributions for valid responses
across time bins for observed RTs (blue) and simulated RTs
generated using the Gamma model (yellow).
effort in models of vigilance.
The proposed model represents an initial step in develop-
ing models of fatigue and vigilance that incorporate directly-
observable neural data. In this model, changes in observed
neural data simply constrain the parameters of the behavioral
model, i.e., a “direct-input approach” (Turner et al., 2017),
implying a unidirectional influence. Future models, however,
will need to simultaneously account for both neural and be-
havioral data and account for the bidirectional relationship
between the two. Similarly, the use of frontal γpower in our
model represents only one potential application of EEG data
in cognitive models; our future research will use similar mod-
els to explore how other neural indices, such as beta (β) and
alpha (α) frequency bands, can be used as observable esti-
mates of fatigue and arousal in computational models of vig-
ilance.
Acknowledgments
The opinions expressed herein are solely those of the authors
and do not necessarily represent the opinions of the United
States Government, the U.S. Department of Defense, the U.S.
Air Force, or any of their subsidiaries, or employees. Distri-
bution A. Approved for public release. Case number AFRL-
2022-1771. The authors thank Bella Veksler and Chris Fisher
for their help in model development and comments on the
paper.
References
Ackerman, P. L. (2011). Cognitive fatigue: Multidisciplinary
perspectives on current research and future applications.
American Psychological Association.
Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S.,
Lebiere, C., and Qin, Y. (2004). An integrated theory of
the mind. Psychological Review, 111(4):1036–1060.
Belavkin, R. V. (2001). Modelling the inverted-U effect in
ACT-R. In Proceedings of the 2001 Fourth International
Conference on Cognitive Modeling, pages 275–76.
Bergum, B. O. and Klein, I. C. (1961). A survey and analysis
of vigilance research. Technical report, No HUMRRO-RR-
8. George Washington University, Alexandria, VA. Human
Resources Research Office.
Bezanson, J., Edelman, A., Karpinski, S., and Shah, V. B.
(2017). Julia: A fresh approach to numerical computing.
SIAM Review, 59(1):65–98.
Borghetti, L., Morris, M., Rhodes, L. J., Haubert, A., and
Veksler, B. (2021). Gamma oscillations index sustained
attention in a brief vigilance task. International Annual
Meeting of the Human Factors and Ergonomics Society.
Borst, J. P. and Anderson, J. R. (2015). The discovery of
processing stages: Analyzing EEG data with hidden semi-
Markov models. NeuroImage, 108:60–73.
Cassenti, D. N., Kerick, S. E., and McDowell, K. (2011).
Observing and modeling cognitive events through event-
related potentials and ACT-R. Cognitive Systems Research,
12(1):56–65.
Delorme, A. and Makeig, S. (2004). EEGLAB: An open
source toolbox for analysis of single-trial EEG dynam-
ics including independent component analysis. Journal of
Neuroscience Methods, 134(1):9–21.
Dinges, D. F. and Powell, J. W. (1985). Microcomputer anal-
yses of performance on a portable, simple visual RT task
during sustained operations. Behavior Research Methods,
Instruments, & Computers, 17(6):652–655.
Doran, S. M., Van Dongen, H. P., and Dinges, D. F. (2001).
Sustained attention performance during sleep deprivation:
evidence of state instability. Archives of Italian Biology:
Neuroscience, 139(3):253–267.
Gunzelmann, G., Gross, J. B., Gluck, K. A., and Dinges, D. F.
(2009). Sleep deprivation and sustained attention perfor-
mance: Integrating mathematical and cognitive modeling.
Cognitive Science, 33(5):880–910.
Gunzelmann, G., Veksler, B. Z., Walsh, M. M., and Gluck,
K. A. (2015). Understanding and predicting the cognitive
effects of sleep loss through simulation. Translational Is-
sues in Psychological Science, 1(1):106–115.
Herrmann, C. S., Fr¨
und, I., and Lenz, D. (2010). Human
gamma-band activity: a review on cognitive and behavioral
correlates and network models. Neuroscience & Biobehav-
ioral Reviews, 34(7):981–992.
Ishii, A., Tanaka, M., and Watanabe, Y. (2014). Neural mech-
anisms of mental fatigue. Reviews in the Neurosciences,
25(4):469–479.
Jongman, G. (1998). How to fatigue ACT-R. In Proceedings
of the second European conference on cognitive modelling,
pages 52–57. Nottingham University Press.
Kass, R. E. and Raftery, A. E. (1995). Bayes factors. Journal
of the American Statistical Association, 90(430):773–795.
Kim, J.-H., Kim, D.-W., and Im, C.-H. (2017). Brain areas re-
sponsible for vigilance: an eeg source imaging study. Brain
Topography, 30(3):343–351.
Lovett, M. C. and Anderson, J. R. (1996). History of suc-
cess and current context in problem solving: Combined
influences on operator selection. Cognitive Psychology,
31(2):168–217.
Mogensen, P. K. and Riseth, A. N. (2018). Optim: A math-
ematical optimization package for Julia. Journal of Open
Source Software, 3(24):615.
Morris, M. B., Haubert, A. R., and Gunzelmann, G. (2020).
Beyond the vigilance end-spurt with event-related poten-
tials. In Proceedings of the Human Factors and Er-
gonomics Society Annual Meeting, volume 64, pages
1258–1262. SAGE Publications Sage CA: Los Angeles,
CA.
Ratcliff, R. (1993). Methods for dealing with reaction time
outliers. Psychological Bulletin, 114(3):510.
Tran, Y., Craig, A., Craig, R., Chai, R., and Nguyen, H.
(2020). The influence of mental fatigue on brain activity:
Evidence from a systematic review with meta-analyses.
Psychophysiology, 57(5):e13554.
Turner, B. M., Forstmann, B. U., Love, B. C., Palmeri, T. J.,
and Van Maanen, L. (2017). Approaches to analysis in
model-based cognitive neuroscience. Journal of Mathe-
matical Psychology, 76:65–79.
Van Dongen, H. (2004). Comparison of mathematical model
predictions to experimental data of fatigue and perfor-
mance. Aviation, Space, and Environmental Medicine,
75(3):A15–A36.
Veksler, B. Z. and Gunzelmann, G. (2018). Functional equiv-
alence of sleep loss and time on task effects in sustained
attention. Cognitive Science, 42(2):600–632.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Vigilance, sometimes referred to as sustained attention, is an important type of human attention as it is closely associated with cognitive activities required in various daily-life situations. Although many researchers have investigated which brain areas control the maintenance of vigilance, findings have been inconsistent. We hypothesized that this inconsistency might be due to the use of different experimental paradigms in the various studies. We found that most of the previous studies used paradigms that included specific cognitive tasks requiring a high cognitive load, which could complicate identification of brain areas associated only with vigilance. To minimize the influence of cognitive processes other than vigilance on the analysis results, we adopted the d2-test of attention, which is a well-known neuropsychological test of attention that does not require high cognitive load, and searched for brain areas at which EEG source activities were temporally correlated with fluctuation of vigilance over a prolonged period of time. EEG experiments conducted with 31 young adults showed that left prefrontal cortex activity was significantly correlated with vigilance variation in the delta, beta1, beta2, and gamma frequency bands, but not the theta and alpha frequency bands. Our study results suggest that the left prefrontal cortex plays a key role in vigilance modulation, and can therefore be used to monitor individual vigilance changes over time or serve as a potential target of noninvasive brain stimulation.
Article
Full-text available
Bridging cultures that have often been distant, Julia combines expertise from the diverse fields of computer science and computational science to create a new approach to numerical computing. Julia is designed to be easy and fast. Julia questions notions generally held as “laws of nature” by practitioners of numerical computing: 1. High-level dynamic programs have to be slow, 2. One must prototype in one language and then rewrite in another language for speed or deploy- ment, and 3. There are parts of a system for the programmer, and other parts best left untouched as they are built by the experts. We introduce the Julia programming language and its design — a dance between specialization and abstraction. Specialization allows for custom treatment. Multiple dispatch, a technique from computer science, picks the right algorithm for the right circumstance. Abstraction, what good computation is really about, recognizes what remains the same after differences are stripped away. Abstractions in mathematics are captured as code through another technique from computer science, generic programming. Julia shows that one can have machine performance without sacrificing human convenience.
Article
The occurrence of mental fatigue during tasks like driving a vehicle increases risk of injury or death. Changes in electroencephalographic (EEG) activity associated with mental fatigue has been frequently studied and considered a promising biomarker of mental fatigue. This is despite differences in methodologies and outcomes in prior research. A systematic review with meta-analyses was conducted to establish the influence of mental fatigue on EEG activity spectral bands, and to determine in which regions fatigue-related EEG spectral changes are likely to occur. A high-yield search strategy identified 21 studies meeting inclusion criteria for investigating the change in EEG spectral activity in non-diseased adults engaged in mentally fatiguing tasks. A medium effect size (using Cohen's g) of 0.68 (95%CI: 0.24-1.13) was found for increase in overall EEG activity following mental fatigue. Further examination of individual EEG spectral bands and regions using network meta-analyses indicated large increases in theta (g = 1.03; 95%CI: 0.79-1.60) and alpha bands (g = 0.85; 95%CI: 0.47-1.43), with small to moderate changes found in delta and beta bands. Central regions of the scalp showed largest change (g = 0.80; 95%CI: 0.46-1.21). Sub-group analyses indicated large increases in theta activity in frontal, central and posterior sites (all g > 1), with moderate changes in alpha activity in central and posterior sites. Findings have implications for fatigue monitoring and countermeasures with support for change in theta activity in frontal, central and posterior sites as a robust biomarker of mental fatigue and change in alpha wave activity considered a second line biomarker to account for individual variability.
Article
Research on sleep loss and vigilance both focus on declines in cognitive performance, but theoretical accounts have developed largely in parallel in these two areas. In addition, computational instantiations of theoretical accounts are rare. The current work uses computational modeling to explore whether the same mechanisms can account for the effects of both sleep loss and time on task on performance. A classic task used in the sleep deprivation literature, the Psychomotor Vigilance Test (PVT), was extended from the typical 10-min duration to 35 min, to make the task similar in duration to traditional vigilance tasks. A computational cognitive model demonstrated that the effects of time on task in the PVT were equivalent to those observed with sleep loss. Subsequently, the same mechanisms were applied to a more traditional vigilance task-the Mackworth Clock Task-providing a good fit to existing data. This supports the hypothesis that these different types of fatigue may produce functionally equivalent declines in performance.
Article
Sleep loss impacts cognitive functioning, and the resulting performance changes can have dramatic consequences in the real world. The increased risk to property and human life has motivated decades of empirical research on fatigue and its effects on performance. Models now exist that can predict the general time course and magnitude of changes in cognitive function caused by fatigue. These models have enabled the development of tools that are useful for shift work and sleep scheduling to improve safety. However, these models are incapable of making a priori predictions regarding the precise, task-specific effects that sleep loss and circadian rhythms will have on performance. Such a capability would make it possible to perform simulation-based risk assessments by conducting systematic evaluations over spaces of system designs, training approaches, policy interventions, and sleep/work schedules. It would also support monitoring technologies to detect behavioral evidence of fatigue. To develop such applications, computational process models that run in simulation are needed to produce behavior predictions in the domains of interest. In this article we review and summarize research committed to precisely this goal, we assess progress to date, and we describe remaining challenges on the path to application. (PsycINFO Database Record (c) 2015 APA, all rights reserved)
Article
Abstract Fatigue is defined as a decline in the ability and efficiency of mental and/or physical activities that is caused by excessive mental and/or physical activities. Fatigue can be classified as physical or mental. Mental fatigue manifests as potentially impaired cognitive function and is one of the most significant causes of accidents in modern society. Recently, it has been shown that the neural mechanisms of mental fatigue related to cognitive task performance are more complex than previously thought and that mental fatigue is not caused only by impaired activity in task-related brain regions. There is accumulating evidence supporting the existence of mental facilitation and inhibition systems. These systems are involved in the neural mechanisms of mental fatigue, modulating the activity of task-related brain regions to regulate cognitive task performance. In this review, we propose a new conceptual model: the dual regulation system of mental fatigue. This model contributes to our understanding of the neural mechanisms of mental fatigue and the regulatory mechanisms of cognitive task performance in the presence of mental fatigue.
Article
Problem solvers often have multiple operators available to them but must select just one to apply. We present three experiments that demonstrate that solvers use at least two sources of information to make operator selections in the building sticks task (BST): information from their past history of using the operators and information from the current context of the problem. Specifically, problem solvers are more likely to use an operator the more successful it has been in the past and the closer it takes the current state to the goal state. These two effects, respectively, represent the learning and performance processes that influence solvers’ operator selections. A computational model of BST problem solving, developed within the ACT-R theory (Anderson, 1993), provides the unifying framework in which both types of processes can be integrated to predict solvers’ selection tendencies.