Access to this full-text is provided by De Gruyter.
Content available from Current Directions in Biomedical Engineering
This content is subject to copyright. Terms and conditions apply.
Current Directions in Biomedical Engineering 2019;5(1):13-16
Open Access. © 2019 Martin Golz, Adolf Schenka, Florian Haselbeck, Martin P. Paulia published by De Gruyter.
Commons Attribution NonCommercial NoDerivatives 4.0 License.
Martin Golz*, Adolf Schenka, Florian Haselbeck and Martin P. Pauli
Inter-individual variability of EEG features
during microsleep events
Abstract: This paper examines the question of how
strongly the spectral properties of the EEG during micro-
sleep differ between individuals. For this purpose, 3859 mi-
crosleep examples were compared with 4044 counterexam-
ples in which drivers were very drowsy but were able to
perform the driving task. Two types of signal features were
compared: logarithmic power spectral densities and en-
tropy measures of wavelets coefficient series. Discriminant
analyses were performed with the following machine learn-
ing methods: support-vector machines, gradient boosting,
learning vector quantization. To the best of our knowledge,
this is the first time that results of the leave-one-subject-out
cross-validation (LOSO CV) for the detection of micro-
sleep are presented. Error rates lower than 5.0 % resulted in
17 subjects and lower than 13 % in another 11 subjects. In
3 individuals, EEG features could not be explained by the
pool of EEG features of all other individuals; for them, de-
tection errors were 15.1 %, 17.1 %, and 27.0 %. In compar-
ison, cross validation by means of repeated random sub-
sampling, in which individuality is not considered, yielded
mean error rates of 5.0 ± 0.5 %.
A subsequent inspection of raw EEG data showed that in
two individuals a bad signal quality due to poor electrode
attachment could be the cause and in one individual a very
unusual behavior, a high and long-lasting eyelid activity
which interfered the recorded EEG in all channels.
Keywords: Microsleep, EEG, periodogram, support-vector
machines, SVM, learning vector quantization, LVQ, gradient
boosting, leave one out, LOSO cross-validation.
https://doi.org/10.1515/cdbme-2019-0004
1 Introduction
The recognition of changes in brain state is a challenging task.
This is especially the case for short-term requirements, e.g.
when microsleep events (MSE) of car drivers should be de-
tected from the EEG. Machine learning methods have been
successfully used for this purpose [1,2]. Because these meth-
ods are based solely on the given data set and on assumptions
about the underlying data generating, stochastic processes,
great care must be taken when selecting data [3]. Therefore,
one research focus should be in validation. On the one hand,
training sample size should be as large as possible in order to
enable confident and accurate learning, and on the other hand
enough test examples should be available, which must be
drawn identically and statistically independent from the same
unknown distribution as the training examples [4].
To the best of our knowledge, this is the first time that re-
sults of a leave-one-subject-out cross-validation (LOSO CV)
[4] for MSE detection are presented. LOSO CV involves train-
ing on data of all subjects except one and validation on data of
the subject held out. This procedure is performed for each in-
dividual such that data of every individual was 𝑛𝐼− 1 times
used for training and at one time used for testing, whereby 𝑛𝐼
is the number of individuals. By holding out data of an indi-
vidual, LOSO CV simulates the case that data of the individual
are not currently available for machine learning training, but
in the future for testing the learned model. It asks how much
the results are influenced if data of an individual are available
or not and therefore, it asks how general the results of machine
learning can be interpreted.
2 Material
LOSO CV analysis is based on data of two studies performed
in our driving simulation lab in the years 2007 [5] and 2016.
The first was completed by 16 (5 ♀, 11 ♂, age: 24.4 ± 3.1) and
the second by 15 young adults (7 ♀, 8 ♂, age: 24.4 ± 2.8). The
procedure was almost the same in both studies: seven driving
sessions with a duration of 40 minutes were started hourly be-
tween 1 and 8 am. All subjects wore an actometer for three
days before the study, so that it could be tested, among other
things, that the end of sleep was at least 8 am and that subjects
initiated sleep latest at 1 am in the two nights before the study.
The study design of both studies ensured that the following
______
*Corresponding author: Martin Golz: University of Applied
Sciences Schmalkalden, Blechhammer 4, Schmalkalden,
Germany, e-mail: m.golz@hs-sm.de
Adolf Schenka, Florian Haselbeck, Martin Patrick Pauli: Uni-
versity of Applied Sciences, Schmalkalden, Germany
https://doi.org/10.1515/cdbme-2019-0004
This work is licensed under the Creative
M. Golz et al., Inter-individual variability of EEG features during microsleep events — 14
four factors were effective for achieving high drowsiness: (1)
Time since sleep was at least 16 hours, (2) Time on task was
relatively long (280 min), (3) Time of day was near the circa-
dian trough, and (4) High monotony occurred due to the driv-
ing task and the absence of communication with others.
There were differences between the both studies, particu-
larly in the technical lab equipment, including the EEG de-
vices. The SigmaPL-Pro® (Neurowerk GmbH, Gelenau, Ger-
many) was used in 2007 and the SomnoScreen® (Somnomed-
ics GmbH, Kist, Germany) in 2016. Electrodes were attached
to eight positions (Fp1, Fp2, C3, C4, O1, O2, A1, A2; refer-
ence electrode: Cz; common average reference).
Based on video recordings of the eye region, head and
shoulders as well as the driving scene, the starting times and
lengths of MSE were determined. MSE have always been de-
termined based on observable behavioral characteristics, in
particular prolonged eyelid closure and slow eye movement.
This visual assessment was performed by a trained person with
long experience in this field [3]. A total of 7903 events con-
sisting of 3859 MSE and of 4044 counterexamples was drawn
from recordings.
3 Methods
The machine learning methods achieved lowest error rates if
the EEG was segmented such that the beginning and end of the
EEG segments were 1 s before and 3 s after the start time of
MSE occurrence. This setting was fixed uniformly for all
channels and all subjects. Details of these empirical optimiza-
tions are addressed in section 4. Logarithmic power spectral
densities (LogPSD) were estimated using the modified peri-
odogram with Hann tapers. Alternatively, the following direct
and indirect PSD estimation methods were tested: Welch,
multi-taper, Yule-Walker, Burg. For them, however, the final
error rates of machine learning were slightly higher. Addition-
ally, feature extraction was extended to the discrete wavelet
transform (DWT) and to wavelet packet transform. Signal de-
composition was performed up to level 8. The total power and
the following entropies were estimated for each detail and for
the approximation coefficient series: Shannon, threshold,
SURE, norm, and mean logarithmic instantaneous power
[6]. The DWT using Daubechies mother wavelet of order 2
(db2) at decomposition level 5 and with the total power as well
as all 5 types of entropies provided feature sets that proved
most successful for machine learning. Thus, from each EEG
segment a 216-dimensional feature vector consisting of 6
measures (total power and 5 entropies) of 6 wavelet coefficient
series (1 approximation, 5 details) in each of 6 EEG channels
(A1 and A2 channels were not processed) was extracted. Com-
parisons of the machine learning performance for these DWT
feature sets and for the LogPSD feature sets will be presented
in section 4.
Empirical optimizations of LogPSD estimates led to the
following setting: spectral bands having a width of 1.0 Hz
across the interval from 0.2 to 40.2 Hz. This way, the number
of LogPSD values was 40 per channel and 240 in total. Pro-
cessing a smaller number of channels led to increased errors.
For supervised training of machine learning methods, a
sample containing pairs of feature vectors and target varia-
bles must be provided:
()
with for LogPSD, for DWT.
As binary target variable the class label was used for
MSE examples and for counterexamples.
The following three different methods of machine learning
have been applied in order to compare their performance:
Learning vector quantization (LVQ)
Gradient boosting (GB)
Support-vector machines (SVM)
LVQ is a neural net consisting of one layer of neurons per-
forming competitive learning. I.e., the neuron whose weight
vector is closest to the current input vector in terms of
a pre-selected vector norm wins the competition among all
neurons and is adapted by the following learning rule:
(2)
with step size
.
The gradient boosting (GB) algorithm creates an ensemble
of decision trees. At each iteration, for each class a new tree is
created such that they correct the errors of the former tree. This
is fulfilled when the squared error loss function has a negative
gradient. Trees added to the ensemble are no more modified.
At the end, for a given data example all trees are retrieved and
a majority decision is made: The class that was calculated by
more trees is selected. Since GB tends to overfit relatively
quickly, the decision trees should be limited according to the
number of branches per level and the depth of the tree. Here
LightGBM was used, a numerically efficient GB variant de-
veloped by Microsoft Inc. [7], which uses existing graphics
processing unit in addition to the central processing unit. How-
ever, our data sets were too small to benefit from using the
graphics processing unit. LightGBM tends to have deeper ra-
ther than wider decision trees [7].
M. Golz et al., Inter-individual variability of EEG features during microsleep events — 15
SVM aims at finding a mapping for any vector
based on all samples . The largest possible
margin of a linear separation function between
the two class domains is sought. I.e. the parameters
must be optimized provided that the distance be-
tween the separation function and the nearest feature vectors,
the support vectors, is maximum. It has been proved that this
problem has a unique solution. The corresponding Lagrangian:
(3)
must be in a saddle point because must be mini-
mized with respect to and maximized with respect to
.
The solution can be formulated explicitly:
(4)
Only the set of support-vectors contribute to equ.
(4). If the training set is not separable, an error term
with different slack variables can be intro-
duced to forgive classification errors (soft margin principle).
This leads to a restriction of the multipliers to the interval
. The regularization parameter C must
be optimized empirically by minimizing mean training errors.
Additionally, the solution should be searched within the high-
dimensional reproducing kernel Hilbert space by applying an
admissible kernel function in order to take advantage
of the blessings of high dimensionality and to get non-linear
separation functions in the input space. This way, the separa-
tion function changes from to
and the equation (3) changes to
. Gaussian kernel functions
were used, which have only
one free parameter to be optimized empirically.
Table 1: Mean and standard deviation of MSE detection errors es-
timated by repeated random subsampling using three different
learning methods. The number of EEG segments and the number
of LogPSD features were equal for all methods.
Method
LVQ
7,903
240
GB
7,903
240
SVM
7,903
240
4 Results
First, results of the repeated random subsampling are shown.
It turned out that the length of EEG segments can be varied
between 4 and 10 s without any degradation in classification
performance. In order to have maximum time resolution the
segmentation length was set to 4 s. The second segmentation
parameter, the time offset between the segment center and the
beginning of the MSE, showed a very sensitive influence on
classification performance (Fig. 1). Only in a small interval the
algorithms can learn accurately; it is optimal to set the center
of the segment 1 s after MSE starts. Also, LogPSD features
resulted in lower error rates than DWT features.
For these optimal settings, the comparison of LVQ, GB,
and SVM learning algorithms resulted in small differences in
mean errors and their standard deviations (Tab. 1). The com-
putational load of SVM was at least 10 times higher than that
of GB and this was about the same factor higher than that of
LVQ. It must be emphasized, however, that these are results
from repeated random subsampling of independent training
and test sets that include data from all subjects in both sets.
Thus, the methods already have information about each sub-
ject during the training. This is not the case for LOSO CV, in
which data of one subject is held out of training and used to
estimate the mean test error. For 31 subjects, the average size
of the training set is and
for
the test set. Since the data of each test person are kept out and
the data of all others are used for training, there are 31 different
mean test errors (Fig. 2). The results show that there are large
differences in learning performance. Data from most subjects
were classified accurately by classifiers who learned from data
from all other subjects. However, if only the results of the best
classifier (SVM) are considered, it turns out that five subjects
are classified with errors higher than 10 %. That is, during
training the methods did not obtain enough information from
data of the other 30 subjects to correctly classify data of the
subject held out. Comparison of learning methods shows that
Figure 1: Mean and standard deviation of classification errors on
the test set versus the time offset between segment center
and MSE start time. Results were estimated using the LVQ
algorithm on LogPSD and on DWT feature sets.
M. Golz et al., Inter-individual variability of EEG features during microsleep events — 16
in LOSO CV the SVM can almost always classify better than
GB and LVQ, often much better.
5 Conclusions
The presented investigation shows that classification of short-
term EEG segments to behavioral characteristics is possible at
low error rates, if the characteristics relate to a change in brain
state, i.e. the microsleep. It has been demonstrated that only in
a short time interval around the onset of MSE an accurate
learnability is given. For segments that are a few seconds be-
fore or after MSE onset, such low errors cannot be achieved,
because obviously no specific brain state can be defined here,
but the normal multi-process mode of the awake brain, which
is limited, however, by high drowsiness.
Direct estimation of LogPSD by the modified periodogram
led to higher accuracies than other estimation methods, which
provide lower variance. This suggests that the trade-off be-
tween bias and variance must be chosen differently for ma-
chine learning methods. They obviously benefit from PSD es-
timation methods with lower bias at higher variance. It re-
mains a future challenge to find a suitable feature extraction
from DWT coefficient series. Power and entropy measures
were of little use in achieving low classification errors in the
following step.
LOSO CV simulates the case that in the future, in a dataset
where learning methods had already been learned in the past,
another set of examples of a new subject will be added. It was
found that data of a few subjects could not be well explained
by data of the other 30 subjects. In order to explain this result,
the raw recordings and the distribution of the extracted fea-
tures were inspected. No easily identifiable cause was found.
One subject (#30) had unusually strong and long-lasting blink
activity in periods of high drowsiness. This activity superim-
posed all channels and may have led to a considerable bias in
estimated spectral features. Others were found to have poor
signal quality, probably due to poor electrode contact resist-
ance in combination with low amplitude EEG.
It might also be that the individually diverse EEG charac-
teristics are a further explanation. This has already been re-
ported from studies on features extracted from 30 s EEG seg-
ments during sleep. In studies with monozygotic and dizygotic
twins, evidence was found of high individuality and high her-
itability of EEG features. One group of authors concluded that
the EEG could possibly be the most inheritable trait of humans
[8]. We are working on carefully processing data from further
studies in our laboratory and incorporating them into LOSO
CV analyses in order to ultimately obtain indications for a con-
clusive explanation.
Author Statement
Authors state no funding involved, as well as no conflict of
interest. Informed consent has been obtained from all individ-
uals included in both studies. Ethical approval: The research
related to human use complies with all relevant national regu-
lations, institutional policies and was performed in accordance
with the tenets of the Helsinki Declaration, and has been ap-
proved by the authors' institutional review board.
References
[1] Golz M, Sommer D. Automatic Recognition of Microsleep
Events. J Biomedizinische Technik 2004;49,Suppl2:332-3.
[2] Peiris M, Jones R, Davidson P, Bones P. Detecting behavior-
ral microsleeps from EEG power spectra. Proc 28th EMBS
Conf 2006:5723-26.
[3] Golz M, Schenka A, Sommer D, et al. The role of expert eva-
luation for microsleep detection. Current Directions in Bio-
medical Engineering 2015;1(1):92–5.
[4] Xu G, Huang JZ. Asymptotic optimality and efficient compu-
tation of the leave-subject-out cross-validation. The Annals of
Statistics 2012;40(6):3003-30.
[5] Golz M, Sommer D, Trutschel U, et al. Evaluation of fatigue
monitoring technologies. J Somnology 2010;14(3):187-99.
[6] Misiti M, Misiti Y, Oppenheim G, Poggi JM. Wavelet toolbox.
The MathWorks Inc., Natick (MA), USA; 2017.
[7] Ke G, Meng Q, Finley T, Wang T, Chen W, et al. LightGBM:
A highly efficient gradient boosting decision tree. In Adv Neu-
ral Inform Proc Syst 2017:3146-54.
[8] De Gennaro L, Marzano C, Fratello F, et al. The EEG finger-
print of sleep is genetically determined: A twin study. Annals
of Neurology 2008,64:455-60.
Figure 2: Results of LOSO cross-validation of all subjects. Mean
classification errors for three different learning algorithms.
Available via license: CC BY 4.0
Content may be subject to copyright.
