CHAPTER 2

LITERATURE REVIEW

2.1 Introduction

The main contribution of this studies is reviewing the research and answer the
following research questions in literature review:
• Common methodologies of collecting and preprocessing EEG data in
drowsiness detection
• Techniques and algorithms use in EEG-based drowsiness detection and their
accuracy.
• Integration of EEG-based drowsiness detection systems with others bio signal
(EMG, ECG) or detection (vehicle-based) and their effectiveness and
reliability.
2.2 Selection of paper

The PRISMA methodology was used to conduct this systematic literature search for
this study. To complete the PRISMA process, there are several steps. First, we must
find papers that are related to our studies from different database, such as Scopus or
Web of Science. Second, removing the duplicated are essential as there are many
articles published in more than one place. Third, we are sorting the remaining articles
and focused on qualifying requirements. For database searches, we selected the
following keywords: Drowsiness Detection EEG. Using this keyword to do a search
between 2022 and 2024 articles. For removing duplicates and not relevant, we
examine the title and abstracts of the research works that found from the database.
The research that are out of criteria are removed. Below are some of the criteria
considered while filtering the not relevant. The result is display on Figure 1.

Inclusion Criteria Exclusion Criteria

Using physiological based (eeg) to Not using physiological based(eeg) to

detect drowsiness detect drowsiness

Written in English Written in other than English

Dataset driver older than 16 Dataset driver younger than 16

Research Article Not Research Article

Figure 2.1 Flow of the process (PRISMA-based) for choosing articles for the final
review.

2.3 Research Objective

2.3.1 Common methodologies of collecting and preprocessing EEG data in

drowsiness detection

There are several different ways to collect EEG signals, but the famous among all are
using electrodes (Alice et al.,2023; Khare and Bajaj,2022; Ahmed et al. ,2023; Hasan
et al.,2024; Hasan et al. ,2022; Chungho and Jinung ,2023; Zhendong et al., 2024;
Elidrissi et al., 2023; Reddy and Behera, 2022; SadeghArefnezhad et al., 2022;
Fouad ,2023). The electrodes will place on the scalp to collect the EEG signals.
Normally these electrodes are using silver/silver chloride electrodes but also some
with gold cup electrodes. Follow 10-20 system (EEG), the electrodes placement
would be placed on specific location such as on the frontal (Fp1, Fp2) and central
(C3, C4) to collect EEG signals (Elidrissi et al., 2023). Signal collected will be
gathered through wearable EEG devices or brain computer interfaces, allowing for
non-invasive monitoring for brain activity (Vasumathi et al., 2023). The electrical
activity of the brain during different states such as drowsy, awake, will all be
recorded and use to detect drowsiness in future step. Next, we require an amplifier to
amplify our EEG signals as most of the EEG signals are very small and require
amplification. There are various EEG amplifiers that are available for EEG such as
OpenBCI, g.USBamp amplifiers, EMOTIV, QUASAR, BioSemi, and Axxonet
(Abhishek et al.,2024).. The quality and performance of the amplifier are important
to ensure the following procedure run smoothly.
After the EEG signal is collected and gathered, it would come to the step of
preprocessing. There are various EEG signal characteristics such as frequency,
amplitude, time domain features and frequency domain features (Abhishek et
al.,2024). Frequency are frequency bands such as delta, theta, alpha, beta and gamma
and each of them associated with specific brain activities. Normally theta would
associate with light sleep and drowsiness. It is also the key indicator to detect
drowsiness. The amplitude of the EEG signals represents the intensity of the
electrical activity in the brain. Time domain features are like root mean square or
standard deviation to provide views from temporal aspect of EEG signals while
frequency domain features like spectral flatness, entropy, and power spectral density
gives information about the frequency distribution of EEG signals. Understanding
EEG signal characteristics is essential to design effective preprocessing steps.
Preprocessing helps to enhance the quality of data before analysis and classification.
There are some common preprocessing techniques using EEG signals such as
filtering, artifact removal, segmentation, and normalization. Filtering techniques such
as bandpass filters are applied to remove noise and unwanted frequencies from EEG
signal. Artifact removal such as Independent Component Analysis (ICA) and wavelet
transform helps to remove components related to eye blinks and muscle movement
and environmental interference. Segmentation helps to divide continuous EEG data
into epochs such as 5 second per segment for analysis while normalization helps to
standardize the EEG data, minimizing the impact of individual discrepancies and
session related variables. By doing so, it guarantees that the subsequent analytical
procedures concentrate on the significant patterns within the brain activity, rather
than affect by irrelevant variations. After the data has been normalized, the features
such as time domain features and frequency domain features will be extracted from
the segmented window and use for classification.

2.3.2 Techniques and algorithms use in EEG-based drowsiness detection and

their accuracy.

There are various techniques and algorithms develop for EEG-based drowsiness
detection, each of them brings with different accuracy and approaches. Table below
shows their algorithm and accuracy.

Reference Algorithm/ Performance Description

Technique

(Sedik et al., FFT, DWT, • 90% for • Use FFT and

multiclass DWT for feature
2023) DNN
• 95% for binary extraction.
class • Use DNN for
classification
(Puspasari et SVM, PCA, • 77.1% for • PCA use for
SVM (Fine feature selection
al.,2023) logistic
gaussian) and
regression • 78% for LR + transformation.
(LR) PCA • LR, SVM use for
• 95.2% for classification
SVM (Medium
gaussian)

(Khare and O-TQWT, • 96.14% for O- • O-TQWT servers

TQWT with for decompose
Bajaj,2022) SVM
SVM EEG signal
• SVM use for
classification
(Chungho LSTM-CNN • 85.6% for • LSTM-CNN for
LSTM-CNN classification
and Jinung
An,2023)

(Alharbey et SVM, CNN • 98% for SVM • SVM and CNN

al.,2023) • 99% for CNN both use for
classification

(Gangadharan SVM, • 78.3% for • SVM use for

SVM classification
and wearable
• Wearable light
Vinod,2022) light weight weight wireless
wireless consumer for less
intrusive towards
consumer driver

(Tarafde et Blinker, • 80.4% for • Blinker algorithm

Decision Tree use for extract
al.,2022) Decision
• 87% for KNN features from
Tree, KNN, • 86.2% for EEG signal for
SVM classification
SVM,
• 86.4% for • Others are all
Ensemble of Ensemble of classification
Bagged Bagged Trees method
• 91.1% for
Trees, AdaBoost
AdaBoost

(Elidrissi et FFT,PSD, • 98.5% for RF • RF for

classification
al., 2023) RF
method
• FFT and PSD
both are
processing
technique
(Bencsik Embedded • Retaining 95% • Embedded
of information Feature Selection
,2023) Feature
needed for DD use for feature
Selection detection selection
Algorithm • Can significantly
reduce set
quantity
(Elmouzoun optimized • 96.4% • Use a
accuracy personalized
Elidrissi et decision tree
• Reduce time machine learning
al., 2022). consumption model
 classification
method

Table 2.1 Summary of previous studies on physiological measures

2.3.3 Integration of EEG-based drowsiness detection systems with others bio

signal (EMG, ECG) or detection (vehicle-based) and their effectiveness and
reliability

The objective of this topic is to evaluate the effectiveness of EEG-based drowsiness

detection system integrate with other measures such as EMG, ECG, or vehicle-based
measure. It helps to understand whether hybrid approach can enhance the accuracy
and dependability or the opposite. A study explores drowsiness detection using
singular and hybrid approaches based on EEG, EOG, and ECG signals (Hasan et
al.,2022). The result has shown that hybrid models outperform compared to single
ones using ANN model. It is because hybrid bio signal-based model reduces metric
disparities between sensitivity and specific metrics, and lead to improve the overall
performance compared to singular bio signal. Other studies have put their
concentration on combining EEG signals with vehicle motion data (Zhendong et
al.,2024). Pearson correlation analysis is performed to find the strongest correlation
combination with EEG signals. The results show 2.53% improvement in EEG
detection and an 8.78% improvement in the accuracy of vehicle motion-based
detection when compared to single measure detection. These shows that multiple
fusion would significantly increase in performance of drowsiness detection. The
fusion has utilized SVM performance after selecting the strongest correlation
combination with Pearson correlation analysis and results in increase detection
accuracy significantly.
2.4 Research Gap

When a driver falls asleep or felt drowsy while driving, they often lose control of the
vehicle and resulting in an accident. In various studies also have shown that driver
drowsiness is one of the main cause of roads accident. If a driver able to receive an
alarm, many accidents could be prevented. Different dataset may bring different
result in accuracy. In most of the studies, it shown that CNN provided the most
accurate result compared to others. It is because CNN are relatively good at image
recognition among the others. It can automatically extract feature from raw input
data such as video or image. It will increase a lot of efficiency and accuracy to
prevent human mistake. Moreover, integrating multiple measures for drowsiness
detection such as EEG and EOG would further improves accuracy and performance.
It is because combination of different measure reduces metric disparities and leads to
a more accurate result.
Developing of CNN provides a promising aspect on drowsiness detection on drivers.
However, it is crucial to recognize the challenges and associated with this approach.
For example, it requires a large amount of labelled data and tends to be slow
compare with another algorithm. It would be a serious issue and require a reliable
solution when it comes to real-time environment. Besides that, there are lack of
research and experiment about hybrid approach between different measures recently
even though is brings a more accurate result most of the time. It is due to time
consumption on finding large amount of data for different measures and integrates
them together.
In future research, the researcher may work on reducing the time and data need for
CNN such as decrease the characteristic of data, the way of preprocessing the data.
Besides that, more real-time environment testing is needed for different algorithm.
Hybrid approach on drowsiness detection should also be listed in consideration to
decrease to rely on specific characteristic and leads to accuracy drops. The intrusive
of the EEG equipment also should takes in count as eventually it needs to be use in
real-time environment. Driver’s comfort would always be the priority if require them
to wear the EEG equipment.