Improving
classification rates for use in fatigue
countermeasure devices using brain
activity
Tran Y, Craig A, Wijesuriya N, Nguyen
H.
Key University Research
Centre in Health Technologies, Faculty of
Engineering
Information Technology,
University of Technology, Sydney,
Australia
Abstract
Fatigue can be defined as a state that
involves psychological and physical tiredness
with a range of symptoms such as tired eyes,
yawning and increased blink rate. It has
major implications for work place and road
safety as well as a negative symptom of many
acute and chronic illnesses. As such there has
been considerable research dedicated to systems
or algorithms that can be used to detect and
monitor the onset of fatigue. This paper
examines using electroencephalography (EEG)
signals to classify fatigue and alert states as
a function of subjective self-report, driving
performance and physiological symptoms. The
results show that EEG classification network for
fatigue improved from 75% to 80% when these
factors are applied, especially when the data is
grouped by subjective self-report of fatigue
with classification accuracy improving to
84.5%.
I. INTRODUCTION
Driver fatigue has major implications for
road safety and is identified as a major cause
of road accidents. It has been estimated to
account for up to 40% of road crashes and
thought to be a result of a decrement in
performance of the driver caused by the
reduction of his/her level of arousal. Fatigue
reduces drivers' ability to react to road and
traffic conditions, resulting in potential
serious injuries and fatalities. Furthermore,
signs such as an increase in driver errors and
slowing of reaction time also suggest that
fatigue impairs cognitive processes. A method of
reducing the risk of fatigue related accidents
is through monitoring or detecting fatigue
changes in drivers through the use of
countermeasure devices.
Since driver fatigue is a complex,
multidimensional phenomenon that gives rise to
behavioral, physiological, and psychological
changes in a person, measures of driver fatigue
have shown to vary. Currently, fatigue
countermeasure methods can be divided into three
categories, and these are: (i)
subjective/psychological measures of fatigue
such as self-report, (ii) performance measures
such as reaction times, steering errors and
deviation in lane position, and (iii)
physiological measures such as eye movement,
heart activity and brain wave activity.
Subjective or self-reported fatigue has
commonly been used in research situations and
found to have moderate to high reliability
between perceived vigilance levels and
performance in a task. However, self-reported
fatigue as a countermeasure method may be
difficult to implement as there are problems
with the validity of self-reported fatigue,
simply because a persons' ability to understand
his/her own fatigue level is believed to be
impaired when they are experiencing fatigue.
Also, subjective reports of fatigue could also
be susceptible to individual bias and
differences and some have suggested that
individual differences influence self-reported
fatigue. Therefore, subjective measures alone
should not be used as a fatigue detection
method.
Driving performance measures are also
potential methods for fatigue countermeasure to
monitor decrements in performance during a
driving task during the onset of fatigue.
Researchers have utilized performance measures
such as reaction time, steering wheel movements,
lane/track position, maintenance of speed, and
driving off the road and lane crossing. One of
the most common changes to occur in driving
performance associated with fatigue is the
change in lateral lane position, leading some to
conclude that steering movement could be a
valuable fatigue decrement measure. In spite of
this, changes in driving behavior can be caused
by reasons other than fatigue, for example,
steering patterns and driving behavior can also
be affected by driver personality and experience
[21]. Therefore this needs to be
clarified before implementing performance
measures such as steering movement as a
countermeasure tool.
Physiological measures such as changes in
eye activity during fatigue have been well
documented. For instance, research has found
that eye blink rate increases after sleep
deprivation as well as during the onset of
fatigue and eye closure rate changes during
fatigue such that alert eye closure during a
blink occurs for around 200 ms (range from 100
to 300 ms) increasing to around 300 ms during
fatigue (range from 200 to 450 ms). Therefore,
eye activity can be considered a good indicator
of fatigue. Another physiological measure that
has been applied to fatigue countermeasure
research is electroencephalography (EEG)
signals, which measures brain activity changes
[10]. Findings include altered alpha
wave activity during impaired cognitive
attention from fatigue and changes in alpha and
theta waves were found to be related to fatigue
and reduced performance [14]. The
results of much of this research suggest that
EEG could be a useful measure of driver fatigue
and more importantly, enables fatigue levels to
be measured directly.
The search for reliable in-vehicle systems
aimed to monitor, detect and alert drivers with
the onset of fatigue has become a major goal of
driver fatigue research. This paper presents
result from using EEG signals to classify the
occurrences of fatigue but as a ftinction of the
participant's subjective response to fatigue,
performance decrements (based on reaction time)
and physiological symptoms (yawning, eye blinks
etc). Given the complex nature of fatigue and
the strong associations shown between the three
factors (subjective, performance and
physiological measure) it was important to
investigate how these factors would influence
the rates of classification on "alert" versus
"fatigue" states in EEG signals.
IV. DISCUSSION
The results show that EEG signals can be
utilized for the classification of fatigue,
making it a promising method for use in fatigue
countermeasure devices. The overall
classification performance of our testing set
was 75.96%.
To test whether classification rates could
be improved the data was then re-grouped as a
function of three factors. These three factors,
subjective self-report, driving performance and
physiological symptoms were chosen as they have
been shown in the literature to be important
factors associated with the onset of fatigue.
After regrouping, the factor that showed the
greatest improvement in classification rates was
subjective self-report. By regrouping into high
and low fatigue groups based on the selfreport
of the participants, the classification of alert
versus fatigue improved up to 84.54%. This
suggests that fatigue most likely contains a
psychological dimension, as suggested by Brown
(1994) who defined it as a "subjectively
experienced disinclination to continue
performing the task at hand". This result shows
that it is important to consider the subjective
view of the person when trying to detect
fatigue.
The other two factors, performance
indicators based on reaction time and
physiological symptoms, also show improvements
to the classification of fatigue. The problem
with using driving performance and reaction time
task as a measure for fatigue in this simulated
study is that slower reaction times may not
indicate fatigue, but perhaps an association in
this case to familiarity with computer games
etc. This factor needs to be tested further.
Grouping based on physiological symptoms did not
improve the classification rates as well as
expected, especially given that physiological
symptoms such as eye blink rate are reported as
good indicators of fatigue. Perhaps this is due
to using EEG signals as the measure to classify
fatigue. This finding supports the two-level
processing model as proposed by Verwey and
Zaidel (2000) which explained a dissociation
between psychological and physiological
indicators of fatigue, suggesting that more
physiological based measures tap into automatic
lower level processing, while the self-report
measures tap into higher level "executive"
cognitive processes and therefore the subjective
self-report grouping used in this sample
improved brain signal classification of fatigue
more than that from physiological symptoms.
