Automatic Drowsiness Detection based on a Single Channel of EEG
Signals using a Hybrid Analysis and Decision Tree Classification
Method under Python
Elmouzoun Elidrissi Mouad
1
a
, Ben Taleb Lhoucine
1
, Lyazidi Aissam
1,2
, Rattal Mourad
1,2
, Mouhsen
Azeddine
1
, Harmouchi Mohammed
1
and Essoukaki Elmaati
1,2
1
Laboratory of Radiation-Matter and Instrumentation, Faculty of Science and Technology, Hassan First University of
Settat, Morocco
2
Higher Institute of Health Sciences, Hassan First University of Settat, Morocco
mharmouchi14}@gmail.com
Keywords: Driver drowsiness, EEG signals, road accident, hybrid system, automatic detection, real-time, machine
learning.
Abstract: The human brain generates millions of signals because they translate all our movements and thoughts, our
physical and psychological state. During driving, all these signals are generated simultaneously. Vigilance
while being behind the wheel is necessary. However, when the roads are monotonous, especially when taking
highways, that alertness state decreases, and the drowsiness state shows off. In Morocco, 1/3 of fatal accidents
on the highway are caused by drowsiness and sleeping while driving. Wherefore, we proposed the idea of
developing an automatic system based on electroencephalogram signals (EEG) that can predict the state of
drowsiness in real-time using several features extracted from the EEG records when this state occurs to drivers
while driving. The proposed work is based on time-frequency analysis of EEG signals based on one single-
channel (FP1-Ref), and drowsiness is predicted using machine learning techniques under Python. The results
are significant where we could reach an accuracy of prediction with an average of 95.7% within 0.062 seconds
using the decision tree classification method.
1 INTRODUCTION
In Morocco, 33.3% is the rate of fatal/deadly
accidents on highways caused by drowsiness and
sleep while driving based on the latest statistics of the
Ministry of Equipment, Transport, Logistics and
Water, Directorate of Roads (Des et al., 2016)
(Statistiques, 2017). These statistics gave us the idea
of developing an automatic system that can predict
drowsiness when occurring and before the situation
becomes worst leading to dangerous accidents.
Therefore, the idea of our system is not new, but
it came to improve the performance and solve the
limitations of the existing ones by using the latest
processing software ‘Python’, also by providing the
best processing techniques ‘time and frequency’ and
machine learning algorithms to perform the most
performant hybrid and automatic method of detecting
drowsiness.
a
https://orcid.org/0000-0002-3784-8967
Though, our work aims to develop a method that
analyses the EEG signals according to time and
frequency domains (Fourier and Spectrum analysis)
based on single channel records before being fed to
our machine learning classifier chosen (DT: Decision
Tree classifier) for the training and testing process.
As a result, our model shows high performance
compared to the existing models, especially the works
that used the same database (Physionet) with an
accuracy of 95.7% and a time consuming of 0.062
seconds.
2 RELATED WORKS
To make our system more efficient, and based on a
heavy analysis, a detailed study was carried on the
existing systems to precise and overcome their
limitations.
Elmouzoun Elidrissi, M., Ben Taleb, L., Aissam, L., Rattal, M., Mouhsen, A., Harmouchi, M. and Elmaati, E.
Automatic Drowsiness Detection based on a Single Channel of EEG Signals using a Hybrid Analysis and Decision Tree Classification Method under Python.
DOI: 10.5220/0010732900003101
In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 301-305
ISBN: 978-989-758-559-3
Copyright
c
2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
301
Detecting drowsiness for drivers was the primary
goal of many researchers. The processing techniques
were wildly variant, where some were based on
sensors only, based on physiological signals like
electroencephalogram, electrocardiogram, electro-
oculogram (EEG, ECG, EOG) or even a mix of these
two techniques. Chang et al, proposed at the year of
2018 a system based on smart glasses that detect
drowsiness based on signals generated by
accelerometers and gyroscopes by capturing the
micro-falls of the head in addition to an infra-red
transceiver that captures the frequency of the blinking
and the degree of closure of the eyes (Chang et al.,
2018).
Other works used algorithms of detecting drowsiness
based on face or eyes recognition like (Ouabida et al.,
2020) & (Dhanalakshmi et al., 2016) or even using
thermal imaging (Kiashari et al., 2020).
However, when we use just signals and
information issued from sensors and no physiological
signals, it is not evident to confirm that the detection
is efficient because blinking or eyes closure or even
the head’s movement are some standard and random
actions that the driver may do while driving, so the
goal is not reached. So, the solution is to use some
techniques based on signals recorded from EEG,
ECG, EOG and others.
To situate our work, the following works used a
single channel study in addition to using the same
dataset of EEG signals available at the Physionet
database (eegmat 1.0.0 and sleep-edfx) to compare
our results and show the improvement added by our
hybrid method. (Belakhdar et al., 2018) proposed a
method of analysing the spectral domain of the EEG
signals under MATLAB by applying Fourier
transform and a classification by artificial neural
network. With this method, their work reached an
accuracy of 88.8%. (Bajaj et al., 2020) reached an
accuracy of 91% after extracting the features using
tunable Q-factor wavelet transform (TQWT) from the
EEG signals, and classifying them using the extreme
machine learning classifier (ELM). The highest
accuracy of 94.45% was reached by (B et al., 2021)
after using a processing method based on wavelet
packet transform (WPT) fed to the extra-tree
classifier.
The proposed work aims to improve the
algorithm’s efficiency of detecting drowsiness of
drivers based on a single channel of EEG signals in
the term of rapidity and accuracy. We will explain the
method with more details in the next section.
3 PROPOSED WORK
Our method proposed in this paper aims to provide a
new hybrid algorithm of detecting drowsiness of
drivers based on temporal and frequential domains by
processing a single channel of EEG records (FP1).
Many researchers confirmed that the FP1 position is
one of the most accurate positions to be based on to
study drowsiness (Strijkstra et al., 2003). The
proposed method is shown in the figure below:
Figure 1: Flowchart of the proposed method.
3.1 Pre-processing Phase
The databases used (eegmat 1.0.0 and sleep-edfx) are
two open databases provided by the Physionet to use
them for researchers works.
All the EEG records were artefact-free. All the
noise was filtered right after the acquisition phase of
these signals due to a 30 Hz low-pass filter and a 50
Hz notch filter to eliminate the noise of the
alimentation source.
The selected subjects were males and females
with a different marge of ages (an average of 18 years
old), and the recording was done under the 10-20
international system.
3.2 Time Segmentation Phase
A segmentation of 3 seconds was applied on the EEG
records in the current work to ensure stationarity of
spectral analysis (Fourier and Power Spectral Density
analysis).
BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)
302
3.3 Features Extraction Phase
This step aims to extract the most significant features
from the single channel of EEG recording following
three domains (temporal, Fourier and spectral).
We designed a function that can extract all the
features one by one, and it scales all of them in the
right shape for the classification procedure.
3.3.1 Temporal Domain Analysis
In the time domain, eight parameters were calculated
to differentiate between the alert/awake and drowsy
states. After calculating the minima, the maxima, the
amplitude peaks and the mean of amplitude peaks,
these parameters are also processed.
The median:
P(y
≤ x) = P(
z
≥ x)
(1)
The mean:
X
̅
=
(2)
The variance:
Var
=
𝛴
(
𝑥𝑖
−
𝑥
)
²
𝑛
(3)
The standard deviation:
Std
=
𝛴
(
𝑥𝑖
−
𝑥
)
²
𝑛
(4)
The root-mean-square:
RMS =
²
(5)
3.3.2 Fourier Domain Analysis
In the present phase, we proposed a frequency
analysis of the EEG signals recorded using Fast
Fourier Transform. After extracting the same feature
as the time domain, we calculated the modulus of
these features to get rid of the imaginary part and to
have just a real significant part.
For 0 ≤ k ≤ N−1:
𝑋
=
𝑥
𝑒
𝑘𝑛
𝑁
(6)
3.3.3 Power Spectral Density Analysis
The spectrum of the signal is calculated using the
burg algorithm to compare the power of each band of
the brain frequencies to have a good differentiation
between the two states of drivers (awake and drowsy).
3.4 Features Selection & Classification
A total of eight machine learning classification
methods were tested in our study to compare the
efficiency and get the best classifier, and secondly, to
select the most appropriate features. As a result, our
hybrid model showed the best accuracies and time
performances.
The classifiers used in our method are Support
Vector Machine (with its four kernels), Gaussian
Process (GP), K-Nearest-Neighbors (KNN),
Multilayer Perceptron (MLP), and Decision Tree
(DT).
3.5 Results
After extracting the features, all the calculated
parameters were scaled and processed using machine
learning classifiers. These classifiers depend on four
parameters which are:
i. True-positive (TP): The subject is drowsy, and
the prediction is positive (Drowsy state predicted)
ii. True-negative (TN): The subject is awake, and
the prediction is negative (Awake state predicted)
iii. False-positive (FP): The subject is awake, and
the prediction is positive (Drowsy state predicted).
iv. False-negative (FN): The subject is drowsy,
and the prediction is negative (Awake state
predicted).
Based on these parameters, the classifier’s
accuracy is calculated according to:
Accuracy =
(7)
Precision =
(8)
Recall =
(9)
F-1 score =
2
∗
∗
(10)
Automatic Drowsiness Detection based on a Single Channel of EEG Signals using a Hybrid Analysis and Decision Tree Classification
Method under Python
303
Table 1: Performance comparison between different
classifiers applied on our selected features
Classifier
First
method
results
Second
method
results
Third
method
results
Hybrid
method
results
SVM
(Linear
kernel)
49.7% 49.9% 49.4% 49.5%
SVM
(Polyno
mial
kernel)
54.6% 85.5% 93.2% 83.6%
SVM
(Sigmoid
kernel)
35.3% 66.8% 88.7% 66.0%
SVM
(RBF
kernel)
71.9% 86.5% 93.3% 87.8%
MLP 49.8% 74.1% 48.9% 75.6%
KNN 90.6% 92.9% 94.1% 93.1%
GP 49.1% 68.9% 49% 56%
DT 49.3% 93.6% 94.3% 95.7%
As we can see, our hybrid model based on the time,
Fourier and PSD domains achieved the best accuracy
compared to all the other selection of features and
classifiers, and here is a comparison in terms of the
executing time, in addition to the confusion matrix
result of the proposed machine learning model
including the processing time.
Our proposed method is the best in terms of both
accuracy of detecting the drowsiness state and the
processing time.
Figure 2: Confusion matrix of the proposed method.
Classifier Accuracy Time (s)
SVM (Linear
kernel)
49.5% 0.983
SVM (Polynomial
kernel)
83.6% 0.722
SVM (Sigmoid
kernel)
66.0% 1.170
SVM (RBF kernel) 87.8% 0.985
MLP 75.6% 5.144
KNN 93.1% 0.142
GP 56% 12.57
DT 95.7% 0.062
BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)
304
4 CONCLUSIONS
The present work proposed a new hybrid method of
detecting drivers’ drowsiness based on time-
frequency analysis of EEG signals from a single
channel (FP1).
We extracted a total of eight features from the
three domains, the time, Fourier and PSD. After that,
we trained eight machine learning models, support
vector machine (with its four kernels), Gaussian
process (GP), K-Nearest-Neighbors (KNN), Multi-
layer Perceptron (MLP), and Decision Tree (DT).
We compared our work to previous ones based on
the same dataset available at the Physionet database,
and the use of a single channel of EEG records. The
added value of our model is the improvement of the
detection’s performance in the term of accuracy,
which achieved 95.7% and the processing time 0.062
seconds.
REFERENCES
B, V. P., & Chinara, S. (2021). Automatic
classification methods for detecting drowsiness
using wavelet packet transform extracted time-
domain features from single-channel EEG
signal. Journal of Neuroscience Methods, 347,
108927. doi: 10.1016/j.jneumeth.2020.108927
Bajaj, V., Taran, S., Khare, S. K., & Sengur, A.
(2020). Feature extraction method for
classification of alertness and drowsiness states
EEG signals. Applied Acoustics, 163, 107224.
doi: 10.1016/j.apacoust.2020.107224
Belakhdar, I., Kaaniche, W., Djemal, R., & Ouni, B.
(2018). Single-channel-based automatic
drowsiness detection architecture with a
reduced number of EEG features.
Microprocessors and Microsystems, 58, 13–23.
doi: 10.1016/j.micpro.2018.02.004
Budak, U., Bajaj, V., Akbulut, Y., Atilla, O., &
Sengur, A. (2019). An Effective Hybrid Model
for EEG-Based. 19(17), 7624–7631.
Chang, W., Chen, L., Member, S., & Chiou, Y.
(2018). Design and Implementation of a
Drowsiness -Fatigue-Detection System Based
on Wearable Smart Glasses to Increase Road
Safety. IEEE Transactions on Consumer
Electronics, PP(c), 1. doi:
10.1109/TCE.2018.2872162
Des, R., & Des, S. (2016). Recueil des statistiques des
accidents corporels de la circulation routière
2016.
Dhanalakshmi, S., Rosepet, J. J., Rosy, G. L., &
Philominal, M. (2016). Drowsy Driver
Identification Using MATLAB. 4(Iv), 198–205.
Kiashari, S. E. H., Nahvi, A., Bakhoda, H.,
Homayounfard, A., & Tashakori, M. (2020).
Evaluation of driver drowsiness using
respiration analysis by thermal imaging on a
driving simulator. Multimedia Tools and
Applications. doi: 10.1007/s11042-020-08696-
x
Ouabida, E., Essadike, A., & Bouzid, A. (2020).
Optik Optical correlator based algorithm for
driver drowsiness detection. Optik -
International Journal for Light and Electron
Optics, 204(December 2019), 164102. doi:
10.1016/j.ijleo.2019.164102
Statistiques, R. D. E. S. (2017). Recueil des
statistiques des accidents corporels de la
circulation routière 2017. 1–129.
Strijkstra, A. M., Beersma, D. G. M., Drayer, B.,
Halbesma, N., & Daan, S. (2003). Subjective
sleepiness correlates negatively with global
alpha ( 8 – 12 Hz ) and positively with central
frontal theta ( 4 – 8 Hz ) frequencies in the
human resting awake electroencephalogram.
340, 17–20. doi: 10.1016/S0304-
3940(03)00033-8
Automatic Drowsiness Detection based on a Single Channel of EEG Signals using a Hybrid Analysis and Decision Tree Classification
Method under Python
305
