Abstract
Introduction
There has been considerable interest
in drowsiness/sleepiness, vigilance and fatigue monitoring in safety-critical
occupational environments. For example, it has been well-recognized that
increased drowsiness, a diminished vigilance level or driving fatigue are
significant risk factors that substantially contribute to the number of motor
vehicle accidents. For this reason, these human performance factors have been
the subject of intensive research studies. However, many more efforts are still
needed to develop monitoring systems that can be successfully applied in real
environment conditions. The main drawback of the existing systems is their
limited capability to deal with a wide range of information sources needed to
cover many aspects influencing human drowsiness, fatigue or vigilance. There is
also a substantial need for the selection of appropriate indicator variables and
measures. The proposed probabilistic framework represents a step towards
building such a monitoring system.
Methods
In the study, an
EEG-based probabilistic model with high-temporal resolution was designed to
predict drowsiness levels of subjects involved in a moving base driving
simulator experiment. Subsequently, the developed drowsiness model was applied
and adapted to the EEG recordings obtained during a vigilance test experiment.
Drowsiness monitoring Forty-four shift workers participated in the
driving experiment. The subjects drove during morning hours directly after a
full night-shift with no sleep. All drivers drove at least 45 minutes with a
maximum of 90 minutes. The collected EEG, EOG and EMG data were used to visually
score the drivers’ drowsiness levels using the Karolinska drowsiness scoring
method. The method was developed for quantification of drowsiness in situations
when the subject is expected to be awake. The method uses EEG and EOG
recordings. EMG is a valuable indicator of arousals and artifacts but is not
included among the criteria for drowsiness. The scoring method is based on the
Rechtschaffen & Kales scoring rules. Slow eye movements, alpha activity
and/or theta activity are considered as signs of sleepiness. The minimum
Karolinska drowsiness score (KDS) value is equal to 0 and is assigned to the
epochs with no sleepiness signs. In contrast, the maximum KDS value KDS equal to
100 is assigned to the epochs showing the sleepiness signs during the whole
20-second epoch. Epochs with the KDS values > 50 are considered to represent
“sleep onset”. The model architecture used is a hierarchical Gaussian mixture
model (hGMM) consisting of two mixtures at the lower hierarchical level. Each
mixture models data density distribution of one of the two drowsiness
cornerstones represented by 4-second long bins/segments belonging to the epochs
with KDS = 0 and KDS > 50, respectively. A compact spectral representation of
the individual segments in the form of autoregressive (AR) model coefficients
was used. The AR coefficients were computed from two unipolar EEG channels
(Fz-A1, Cz-A2) and a bipolar EEG channel (Oz-Pz). The AR coefficients from all
three EEG channels were concatenated into a 30-dimensional input vector. The
measure of different drowsiness states of drivers are the visually assigned KDS
values. Instead of the strict following of these values, the hGMM is initialized
using the information about the extreme drowsiness states–low and high
drowsiness cornerstones–only. An output of the model is the continuous curve of
the posterior probabilities reflecting a belief about class-membership of the
particular data input to the one of the drowsiness cornerstones.
Vigilance monitoring Fifteen subjects completed the ‘Quatember-Maly’
vigilance test during three separated 50 minutes long sessions. The test is a
computerized version of the Mackworth clock test, which was developed for
vigilance measurements. On the screen a bright spot, moving in tiny jumps, can
be seen describing a circular path. Occasionally the spot makes a “double jump”
to which the test subject is required to respond. The hGMM architecture
developed on driving simulator data was adapted to the “clock test” data using
the following strategy. The drowsiness model was applied to the new data first.
The posterior values of each 4-second data segment were computed and the
segments with the posterior values lower than 0.2 or greater than 0.8 were
extracted. The selected segments were subsequently used to re-train the model.
Elements of the two Gaussian mixtures, for which a prior value was reduced to
zero, were removed and the reduced model was re-trained until all priors were
stabilized. After completing the training step all clock test data were applied
to the model. The obtained posterior values were analyzed and compared with the
reaction time values.
Results
Drowsiness monitoring
First, the hGMM was trained to correctly classify 4-second segments belonging to
the two driver’s drowsiness cornerstones. On average 78% correct classification
rate was achieved making about the same number of errors on each of the two
classes. Second, the hGMM was used to reflect continuous changes in drowsiness
levels. The posterior values from all data segments arranged in the course of
the driving experiment were computed. The Spearman rank correlation coefficient
was computed between smoothed curves of the predicted levels of drowsiness
(posterior values) and the KDS values. The rank correlation coefficient among
the subjects varied in the range of 0.2 - 0.8 with the mean value equal to 0.44.
Good visual agreement between the smoothed curves was observed in the majority
of subjects.
Vigilance monitoring Expect for a few cases, small
values of the Spearman correlation coefficient computed between reaction times
and predicted vigilance were observed on data colleted during two morning time
test sessions. The same is true for the mean values averaged over all subjects.
This is in contrast to the results obtained from the afternoon time session. In
this case, the vigilance levels and reaction times, averaged over all subjects,
showed increasing trend over the course of 50 minutes. The rank correlation was
equal to 0.82 indicating a high agreement between the vigilance and reaction
time curves.
Discussion
The study has shown promising results in
applying the proposed probabilistic framework to two drowsiness and vigilance
level monitoring tasks. The computations associated with the approach are fast
enough to build up a practical real-time system. Moreover, such system can be
applied and tested in different human drowsiness, vigilance, or fatigue
monitoring problems. Future steps will involve the principled probabilistic
inclusion of other physiological signals, visual measures of drowsiness and
contextual information into the system.
Go back