Electroencephalography Analysis Frameworks for the Driver Fatigue

Problem: A Benchmarking Study

Kemalcan Kucuk, Efe Ismet Yurteri and Beren Semiz

Department of Electrical and Electronics Engineering, Koc¸ University, Istanbul, Turkey

Keywords:

Driver Fatigue, Electroencephalography, Channel Selection, Feature Extraction, Classiﬁcation.

Abstract:

Driver fatigue problem is a major factor contributing to trafﬁc accidents globally, making its analysis and

detection crucial for early prevention. Among various approaches for detecting driver fatigue, electroen-

cephalography (EEG) processing is one of the most widely employed techniques. This study investigates

different feature extraction and machine learning methodologies for detecting driver fatigue using EEG sig-

nals, and provides a comparative performance analysis against existing methods. To that aim, we used a

publicly available dataset collected during a simulated driving task and applied our feature extraction methods

to the concurrently recorded EEG signals. Various features from distinct groups were extracted to serve as

the foundation for subsequent analyses. The 30 channels from the original dataset were individually evaluated

based on the performance of machine learning algorithms trained on each channel, allowing for the selection

of the four most optimal channels. Using these selected channels, the different subsets of extracted features

were then compared based on their accuracy values. For further analysis, the features were ranked using both

ANOVA and Chi-Squared feature selection methods to examine the impact of the number of features on model

performance. Each model was ﬁrst trained using a standard training-testing split, where the highest-scoring

model was a Support Vector Machine (SVM) achieving a test accuracy of 90.73%. Additionally, using a

Leave-One-Out Cross-Validation (LOOCV) approach, the highest performing model was found to be the k-

Nearest Neighbors (K-NN) classiﬁer with an average test accuracy of 70.45%. The analyses and comparisons

presented in this study may serve as a basis for developing real-time applications and for further in-depth

investigations.

1 INTRODUCTION

Driver fatigue is commonly characterized by a state of

reduced physical and cognitive alertness and is recog-

nized as a major contributor to trafﬁc accidents. It

impairs the individual’s ability to accurately perceive

and respond to stimuli, thereby compromising driv-

ing performance and increasing the risk of accidents

(Lal and Craig, 2001; May and Baldwin, 2009; Con-

nor et al., 2002; Philip et al., 2005). The condition is

ampliﬁed by various factors such as prolonged driv-

ing durations, inadequate rest, and monotonous road

conditions (May and Baldwin, 2009). In addition to

the countermeasures taken by the drivers, efforts to

mitigate the risks associated with driver fatigue span

across multiple disciplines. These approaches range

from vehicle and road monitoring systems—such as

collision avoidance and lane departure warning sys-

tems—to driver monitoring systems that utilize vari-

https://orcid.org/0000-0002-7544-5974

ous physiological signals to infer the driver’s level of

fatigue.

The most common types of physiology-based de-

tection mechanisms include heart rate monitoring,

blood pressure monitoring, electroencephalography

(EEG) processing, facial landmarking, eye-tracking,

head-posture estimation and respiration estimation

emcov

a et al., 2021). These detection methods typ-

ically rely on preprocessing the signal data via differ-

ent techniques, extracting features from the data, and

utilizing varying modelling techniques to make infer-

ences about the driver. Recently, machine learning

and deep learning models have become predominant

among these detection methods, as they offer robust

and adaptable pipelines that can be applied across di-

verse scenarios.

Among the aforementioned methods, EEG sig-

nals have been utilized in various applications as they

directly reﬂect the brain activity and functionality.

These applications range from diagnosing neurode-

Kucuk, K., Yurteri, E. I. and Semiz, B.

Electroencephalography Analysis Frameworks for the Driver Fatigue Problem: A Benchmarking Study.

DOI: 10.5220/0013086900003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 829-836

ISBN: 978-989-758-731-3; ISSN: 2184-4305

829

generative diseases to developing brain-computer in-

terfaces (BCIs) (Singh and Krishnan, 2023). Clinical

neuroscience heavily relies on the assessment and di-

agnosis of brain disorders using EEG signal analysis.

The purpose of EEG is to identify irregular brain ac-

tivity patterns linked to a range of neurological con-

ditions, including epilepsy, Alzheimer’s disease, and

sleep disorders. By analyzing EEG signals, clinicians

can detect distinct patterns—such as the slowing of

brain waves or the presence of epileptic spikes—that

are indicative of speciﬁc neurological disorders. Ad-

ditionally, EEG is valuable for monitoring disease

progression and evaluating the effectiveness of ther-

apeutic interventions.

EEG signals are increasingly employed in every-

day contexts to monitor cognitive states, particularly

attention levels, in diverse scenarios such as educa-

tional settings and driving environments. In driv-

ing, EEG provides real-time neurophysiological feed-

back on a driver’s cognitive load, enabling the de-

tection of mental states associated with distraction

or fatigue. This capability is crucial for developing

driver-assistance systems that can issue alerts when

focus diminishes, thereby potentially reducing the

risk of accidents. Furthermore, EEG-based assess-

ments in educational environments facilitate the mon-

itoring of students’ attention and cognitive engage-

ment during instructional activities, offering educa-

tors insights into cognitive workload and engagement

levels. Researchers and practitioners can develop in-

terventions to improve attention processes in a variety

of real-world settings and gain important insights into

attention dynamics by integrating EEG into everyday

environments.

This research aims to evaluate various feature ex-

traction techniques, assess the signiﬁcance of optimal

channel selection, and compare the efﬁcacy of differ-

ent machine learning algorithms to develop a more

robust and accurate framework for fatigue detection.

By systematically analyzing the contributions of each

feature extraction method and reﬁning channel selec-

tion, the study seeks to improve model performance

and reliability in detecting cognitive fatigue.

2 METHODS

2.1 Dataset Description

A publicly available EEG dataset was used to com-

pare channel performances and feature extraction

methods. The dataset is based on Cao et al.’s work

(Cao et al., 2019), where subjects (age: 22-28 years)

in a simulated driving task were required to keep the

vehicle in a straight line as random deviations were

introduced to the vehicle’s trajectory. EEG signals

(at 500 Hz sampling rate) consisting of 30 channels

for each subject were recorded in the task. Cao et al.

also published a preprocessed version of the dataset

where the raw signals were ﬁltered with 1 Hz high-

pass and 50 Hz low-pass FIR ﬁlters, and ocular and

muscular artifacts were removed. Cui et al. (Cui

et al., 2022) published a labeled version of Cao et

al.’s work in which they extracted 3-second long EEG

samples (downsampled to 128 Hz) prior to onset de-

viation, and these samples were labeled as alert or

drowsy, based on the subjects’ reaction time as it was

described in (Wei et al., 2018). After labeling, Cui et

al. did additional processing to keep the balance in

class labels: In the end, their dataset had 2022 sam-

ples (1011 alert samples, 1011 drowsy samples) from

11 different subjects, and this dataset was chosen for

the purposes of this work.

2.2 Feature Extraction

We extracted features from two main domains: time-

domain features and frequency-domain features. Ex-

traction was done on each of the 30 channels of the

3-second long EEG signals, resulting in 60660 sam-

ples. In total, 49 features were extracted from each

sample, and this 60660 by 49 dataset was set as the

baseline for selecting different subsets of the dataset

going forward.

2.2.1 Time-Domain Features

Time-domain features are less commonly used in

EEG processing with respect to the other feature do-

mains in the literature. However, they can still provide

meaningful information about the data, which we will

compare against frequency domain features.

Statistical Features: To assess the trends seen in

the signals, the following statistical features were ex-

tracted from the EEG signal samples: zero crossing

rate, skewness, kurtosis, mean absolute difference,

root mean square, the exponent factor of the Hurst

exponent obtained via detrended ﬂuctuation analysis

(Jenke et al., 2014; Wang et al., 2022).

Hjorth Parameters: In addition to the temporal

features, Hjorth parameters were also investigated. In

the original work by Hjorth, activity, mobility and

complexity were developed (Hjorth, 1970).

The activity of a signal is essentially its variance

and it can contain information about signal power.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

830

The mobility of the signal quantiﬁes the standard de-

viation of the signal’s ﬁrst derivative relative to the

standard deviation of the original signal, and it can be

used as an indicator for rapid changes observed in the

signal-of-interest. The complexity refers to the stan-

dard deviation of the signal’s second derivative rela-

tive to the standard deviation of the ﬁrst derivative,

normalized by the mobility. Overall, complexity rep-

resents any signal’s deviation from a pure sine wave,

and can be used to assess the level of signal complex-

ity.

2.2.2 Frequency-Domain Features

Analysis and classiﬁcation of EEG signals commonly

include utilization of frequency-domain features. One

of the most popular approaches is to decompose a sig-

nal into different frequency bands, as the respective

bands can carry distinct information about the sub-

ject’s cognitive level (Lal and Craig, 2002; Simon

et al., 2011).

Band Powers: For the extraction of band powers,

the EEG signals were decomposed into gamma (30-

64 Hz), beta (13-30 Hz), alpha (8-13 Hz), theta (4-8

Hz), and delta (1-4 Hz) bands using Welch’s method

on estimated Power Spectral Density. The power of

each frequency band was then stored as an individ-

ual feature. Along with the powers, the ratio be-

tween each band power was also extracted, as the

work by Minhas et al. showed that the increase of the

theta/alpha ratio can indicate increased drowsiness or

fatigue of a driver (Minhas et al., 2024a; Minhas et al.,

2024b).

Spectral Features: In addition to the commonly

used EEG features, we have also incorporated the fol-

lowing features used in speech and audio processing

applications to check whether they could provide use-

ful characterizations in the driver fatigue task:

• Spectral Centroid

• Spectral Roll-Off

• The First 10 Mel-Frequency Cepstral Coefﬁcients

(MFCCs)

• The Mean of the First 6 Mel-Spectrogram Seg-

ments

• The Energy of the First 6 Mel-Spectrogram Seg-

ments

For the last two types of features, the Mel-

Spectrogram of each EEG signal was extracted, and

then the spectrogram was divided into 32 different

horizontal segments, i.e., into different frequency

bins. Of these segments, only the initial segments

were containing the majority of the information for

the spectrogram, and preliminary testing with differ-

ent segments yielded lesser performance. Thus, the

ﬁrst six segments were chosen to be used in the anal-

ysis.

2.3 Channel Selection

With the initial features extracted from all 30 chan-

nels, we have proceeded to select target channels

to perform feature analysis. This was done both to

reduce the computational load and to eliminate the

channels that can have irrelevant information to be

able to train the models with more concise and rel-

evant data.

For testing the channels, the following procedure

was followed: The data from a given channel was

prepared for model training with a 80/20% train-

ing/testing split. A K-Nearest Neighbor (K-NN) Clas-

siﬁer and an Extreme Gradient Boosting Classiﬁer

(XGB) were trained on the data for the given chan-

nel, and later, each channel was ranked with respect

to their test accuracies. The K-NN model was cho-

sen due to its ability to provide an overview of chan-

nel performances while maintaining a rather low com-

putational cost for training (Hu, 2017)(Dreißig et al.,

2020). The XGB model was chosen since it is one

of the most well-performing machine learning models

for EEG classiﬁcation tasks (Choi et al., 2018)(Parui

et al., 2019). While this procedure does not allow in-

depth performance analysis for a given channel, it can

help us discern channels and test channel combina-

tions.

Table 1: XGB Performances of the Top Ranking Channel.

Cz CP3 CPz P3

Test Accuracy 81.23 81.48 80.99 80.00

Sensitivity 77.16 75.13 77.66 74.62

Precision 83.06 85.06 82.26 82.58

F1-Score 80.00 79.78 79.90 78.40

AUC 81.13 81.31 80.90 79.86

We then tested the top-ranking channels with in-

crementing combinations, i.e., we ﬁrst tested the top

2 channels combined, then the top 3, and so on. The

best-performing trial was the one leveraging the com-

bination of the top 4 channels, which were Cz, CP3,

CPz, and P3, as can be seen from Table 1.

Electroencephalography Analysis Frameworks for the Driver Fatigue Problem: A Benchmarking Study

831

2.4 Feature Analysis

To have a better understanding of the extracted fea-

tures and the dataset as a whole, we tested the per-

formances of feature subsets previously described.

For testing different feature subsets in terms of test

accuracy, a similar setup to channel testing was used

with 80/20% training/testing split for the models in

Table 2.

Table 2: Feature Subset Performances.

K-NN XGB

Statistical 71.26 72.56

Hjorth 73.61 72.62

Time 73.00 77.56

Band Powers 71.08 74.72

Spectral Features 67.31 64.71

Frequency Features 75.28 83.25

Aside from the performance of frequency fea-

tures, which could be attributed to the fact that the

subset had more features, there were minor differ-

ences between the subsets, and no particular feature

subset was dominantly performing across the mod-

els. Moving on from assessing subsets to determine

the signiﬁcance of a given feature, two feature analy-

sis methods were utilized: ANOVA and Chi-Squared.

After the scores for both methods were computed, all

features were sorted by decreasing scores for further

analysis.

2.5 Model Selection

Table 3: Model Performances.

Test Acc. Recall Precision F1 AUC

K-NN 74.78 75.53 74.24 74.88 74.79

SVM 90.73 93.04 88.85 90.90 90.74

DTC 76.95 75.16 77.76 76.44 76.94

RFC 83.93 85.71 82.63 84.15 83.94

LR 75.11 76.15 75.31 75.73 75.71

NN 83.44 87.33 80.90 83.99 83.46

XGB 86.53 88.07 85.32 86.67 86.65

Using all of the extracted features from the selected

channels (Cz, CP3, CPz, P3), the models in Table 3

were trained on the 80/20% splits. The K value for

the K-NN model was chosen as the square root of the

sample size based on the work done by Hassanat et

al., which showed that the square root value provides

sensitivity to noise and overall better generalization

performance. (Abu Alfeilat et al., 2019). The hyper-

parameters of the SVM and XGB models were chosen

based on the result of a Grid Search Cross Validation

algorithm (Kumar et al., 2016).

2.6 Leave-One-Out Cross Validation

In biomedical signal processing applications using

machine learning, it is common to see the use of

leave-one-out cross validation (LOOCV) algorithms

to ensure that the models have overall better general-

ization, which can lead to less susceptibility to inter-

subject variability (Kunjan et al., 2021). Again, using

all of the extracted features from the selected chan-

nels, we implemented a LOOCV algorithm as it had

the previously discussed beneﬁts compared to a reg-

ular training-test split. The selected models for the

LOOCV were the same models in the regular train-

ing/test split to make comparisons between their per-

formances.

3 RESULTS AND DISCUSSION

3.1 Feature Performances

The results for the rankings of top-performing fea-

tures can be seen in Table 4. The rankings gener-

ated by both the ANOVA and Chi-Squared feature se-

lection methods are largely similar, with the highest-

performing features appearing in roughly the same

positions, although there are minor variations. This

suggests that both methods prioritize similar features

for model performance, despite slight differences in

their respective rankings.

Following the descending order from Table 4, the

test accuracy of each model was plotted against the

number of features in Figure 1, with the feature count

increasing incrementally based on this sorted rank-

ing. The decision to use a decreasing order was based

on the assumption that higher-ranked features would

contribute the most to the model’s performance, en-

abling early gains in accuracy. As the remaining,

lower-ranked features are added, the expectation was

to observe a plateau in performance, indicating the

point at which additional features no longer signiﬁ-

cantly enhance model accuracy.

From Fig. 1, it can be seen that around the 30-

feature mark, most of the model performance either

began to stagnate or ﬂuctuated between a relatively

low percentile range. This conﬁrms our ﬁndings in

the preliminary analysis, which suggested that in-

creasing the number of features does not necessarily

result in higher model performance. Given this and

the fact that the feature amount is already relatively

low for most algorithms, we decided not to implement

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

832

Table 4: Chi-Squared and ANOVA Feature Rankings.

Rank Chi-Squared Feature ANOVA Feature Rank Chi-Squared Feature ANOVA Feature

1 Complexity Energy 3

Mel Band 11 Energy 2

Mel Band Energy 6

Mel Band

2 Alpha Power Mean 3

Mel Band 12 Mean 6

Mel Band Mean 6

Mel Band

3 Energy 5

Mel Band Mean 4

Mel Band 13 Spectral Centroid Root Mean Square

4 Energy 3

Mel Band Energy 2

Mel Band 14 Gamma Pow / Beta Pow Spectral Centroid

5 Energy 6

Mel Band Energy 4

Mel Band 15 Mean 2

Mel Band Alpha Pow

6 Spectral Roll-Off Energy 5

Mel Band 16 Theta Pow Theta Pow

7 Mean 5

Mel Band Mean 2

Mel Band 17 Root Mean Square 9

MFCC

8 Energy 4

Mel Band Mean 5

Mel Band 18 Beta Pow / Alpha Pow 3

MFCC

9 Mean 3

Mel Band Spectral Roll-Off 19 Gamma Pow / Alpha Pow Gamma Pow / Beta Pow

10 Mean 4

Mel Band Complexity 20 Alpha Pow / Delta Pow Mean 1

Mel Band

Figure 1: Figure of the model performances with an incre-

mental feature subset according to the ANOVA sorted fea-

ture list.

any feature selection methods based on the result of

ANOVA and the Chi-Squared analysis.

As a sample plot, the mean of the 2nd mel-band

and complexity are plotted in Fig. 2, where alert

and drowsy states are represented in red and blue col-

ors, respectively. The center of mass for each state is

shown with square markers. The features are arbitrar-

ily chosen from the top 20 ranking features according

to the sorted list represented in Table 4, and the scatter

plot can be seen as an indication of the intertwined na-

ture between the different states. Even still, the center

of mass for each state is distinct enough, which makes

a suitable ground for inferences.

The results show that Mel-spectrogram features,

spectral roll-off and complexity from Hjorth fea-

tures deemed effective. For the features related to

frequency bands, features involving the alpha band

were prominent, as discussed in the previous sections.

Along with the alpha band, features involving the

delta, beta, and gamma bands were also high-scoring

features, in alignment with (Minhas et al., 2024b).

Figure 2: Scatter plot of the complexity vs. the mean of the

2nd Mel-band.

3.2 Model Performances

The SVM and XGB models consistently emerged

as top performers across various evaluation metrics.

However, it is important to note that these models

were speciﬁcally ﬁne-tuned to this dataset. As a

result, their performance might not generalize well

to other datasets, potentially leading to suboptimal

results when applied to different data distributions.

This highlights the need for further testing and val-

idation on diverse datasets to ensure the robustness

and adaptability of these models beyond the current

study. However, these models can also serve as ef-

ﬁcient candidates for subject-speciﬁc classiﬁcation

tasks where prediction accuracy takes precedence. In

scenarios where capturing individual differences and

intra-subject variability is crucial, SVM and XGB can

Electroencephalography Analysis Frameworks for the Driver Fatigue Problem: A Benchmarking Study

833

Table 5: Leave-One-Subject-Out Cross Validation Performances.

Subject ID K-NN SVM DTC RFC LR NN XGB Mean

1 79.65 80.45 77.26 81.12 84.57 78.72 78.59 80.05

2 55.11 57.01 56.44 57.39 61.36 57.58 57.95 57.55

3 63.83 64.17 67.67 68.33 69.50 50.17 68.33 64.57

4 71.79 67.74 59.63 61.32 57.60 57.43 61.66 62.45

5 67.75 69.87 67.08 67.75 62.05 73.33 67.75 67.94

6 79.22 73.34 70.78 62.50 72.44 50.45 71.08 68.54

7 58.09 62.01 64.95 68.00 66.91 53.68 63.24 62.41

8 67.14 67.05 64.20 68.00 68.94 51.61 67.61

64.94

9 86.15 85.67 77.00 83.28 84.16 69.38 80.73 80.91

10 80.09 77.31 78.93 75.23 72.92 78.47 72.92 76.55

11 66.15 65.71 67.59 67.92 68.47 50.00 65.27 64.44

Mean 70.45 69.80 68.15 69.49 69.90 60.98 69.16 68.65

deliver effective results. Their ability to adapt to per-

sonalized data makes them particularly suitable when

generalizing across subjects is not the primary goal,

but rather the ﬁne-tuning of models to individual char-

acteristics to optimize accuracy is prioritized.

3.3 Leave-One-Out Cross Validation

(LOOCV) Accuracy

Table 5 shows the test accuracy of each model with

their respective subject and model means. The means

of the overall subject performances vary on a rela-

tively large range, which can be caused by many fac-

tors such as the method of experimentation, any ag-

gravating movement during measurement, noise and

inconsistencies in the sensor data.

The highest-performing subject was subject 9,

with the highest overall performing model being the

SVM (as well as being the highest-performing model

in the previous section). Whereas the lowest-scoring

subject was subject 2, and the lowest-scoring model

on average was the Decision Tree Classiﬁer. As ex-

pected, the mean performance of the models trained

using LOOCV was lower compared to the results

from the previous section. This outcome can be at-

tributed to the fact that some models were ﬁne-tuned

for speciﬁc data splits, leading to reduced general-

ization capability when applied to new, unseen sam-

ples. Despite the lower performance, LOOCV results

may still be appropriate for applications that prioritize

inter-subject generalization, as LOOCV rigorously

tests the model’s ability to generalize across different

individuals. Such results can be valuable for imple-

mentations where accommodating variations between

subjects is critical.

4 CONCLUSION

In this paper, we investigated various feature ex-

traction techniques and machine learning models ap-

plied to EEG signals for the detection of driver fa-

tigue. Leveraging the extracted features, we con-

ducted a comparative analysis of different EEG chan-

nels within the dataset to identify the most relevant

channels for this application. By doing so, we aimed

to determine which channels contribute most signif-

icantly to the detection of fatigue, enabling a more

targeted and effective approach to improving model

performance for this speciﬁc use case.

Using the selected EEG channels, we evaluated

various feature subsets and analyzed their perfor-

mance across different machine learning algorithms.

These comparisons highlighted the most impactful

features and frequency bands for the driver fatigue

detection task. Additionally, we employed feature

ranking methods to assess the individual contribu-

tion of each feature, providing insight into their rel-

ative effectiveness for solving the problem. The ma-

chine learning models demonstrated promising per-

formance depending on the context; both regular

train-test split and LOOCV approaches were applied

to the dataset, and their respective results were dis-

cussed to assess the robustness and generalizability

of the models for different use cases.

Our ﬁndings on the channel analysis and feature

extraction methods align with the existing literature,

and they show the availability of different methods

for practical applications. The methods gathered from

this research can be quite swiftly translated into a real-

time embedded application for a driver fatigue detec-

tion system. Our contribution to the literature is cen-

tered on a focus towards high-yield, low-cost feature

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

834

and channel analysis for EEG-based driver fatigue de-

tection. We also implemented different feature ex-

traction methods from different disciplines (such as

speech processing problems), which are not com-

monly used in EEG processing, and they returned

comparable results to what was currently available

in the literature. These unconventional approaches

yielded results that were comparable to existing meth-

ods in the ﬁeld, demonstrating their potential to en-

hance EEG analysis without adding signiﬁcant com-

putational or ﬁnancial overhead.

This work broadens the scope of feature extraction

in EEG studies by incorporating diverse methodolo-

gies, while maintaining a focus on practical and cost-

effective solutions. To further extend this research, a

key focus for future work could be the development of

real-time applications based on these ﬁndings. These

real-time systems could provide instant feedback to

drivers, improving road safety by mitigating the risk

of fatigue-induced accidents. Additionally, some of

the available features and extraction methods were

preliminarily excluded from the scope of this paper

due to their high computational cost and low impact

on performance. However, conducting more compre-

hensive studies to evaluate the viability of these meth-

ods in various EEG applications could prove beneﬁ-

cial for the literature.

REFERENCES

Abu Alfeilat, H. A., Hassanat, A. B., Lasassmeh, O.,

Tarawneh, A. S., Alhasanat, M. B., Eyal Salman,

H. S., and Prasath, V. S. (2019). Effects of distance

measure choice on k-nearest neighbor classiﬁer per-

formance: A review. Big Data, 7(4):221–248.

Cao, Z., Chuang, C.-H., King, J.-K., and Lin, C.-T. (2019).

Multi-channel eeg recordings during a sustained-

attention driving task. Scientiﬁc Data, 6(1).

Choi, H.-S., Kim, S., Oh, J. E., Yoon, J. E., Park, J. A.,

Yun, C.-H., and Yoon, S. (2018). Xgboost-based

instantaneous drowsiness detection framework using

multitaper spectral information of electroencephalog-

raphy. In Proceedings of the 2018 ACM International

Conference on Bioinformatics, Computational Biol-

ogy, and Health Informatics, pages 111–121.

Connor, J., Norton, R., Ameratunga, S., Robinson, E.,

Civil, I., Dunn, R., Bailey, J., and Jackson, R. (2002).

Driver sleepiness and risk of serious injury to car oc-

cupants: population based case control study. BMJ,

324(7346):1125.

Cui, J., Lan, Z., Liu, Y., Li, R., Li, F., Sourina,

O., and M

uller-Wittig, W. (2022). A compact

and interpretable convolutional neural network for

cross-subject driver drowsiness detection from single-

channel eeg. Methods, 202:173–184. Machine Learn-

ing Methods for Bio-Medical Image and Signal Pro-

cessing: Recent Advances.

Dreißig, M., Baccour, M. H., Sch

ack, T., and Kasneci, E.

(2020). Driver drowsiness classiﬁcation based on eye

blink and head movement features using the k-nn al-

gorithm. In 2020 IEEE Symposium Series on Compu-

tational Intelligence (SSCI), pages 889–896. IEEE.

Hjorth, B. (1970). Eeg analysis based on time domain prop-

erties. Electroencephalography and Clinical Neuro-

physiology, 29(3):306–310.

Hu, J. (2017). Comparison of different features and classi-

ﬁers for driver fatigue detection based on a single eeg

channel. Computational and Mathematical Methods

in Medicine, 2017:1–9.

Jenke, R., Peer, A., and Buss, M. (2014). Feature extraction

and selection for emotion recognition from eeg. IEEE

Transactions on Affective Computing, 5(3):327–339.

Kumar, N., Khaund, K., and Hazarika, S. M. (2016). Bis-

pectral analysis of eeg for emotion recognition. Pro-

cedia Computer Science, 84:31–35.

Kunjan, S., Grummett, T. S., Pope, K. J., Powers, D. M. W.,

Fitzgibbon, S. P., Bastiampillai, T., Battersby, M., and

Lewis, T. W. (2021). The necessity of leave one sub-

ject out (loso) cross validation for eeg disease diagno-

sis. In Mahmud, M., Kaiser, M. S., Vassanelli, S., Dai,

Q., and Zhong, N., editors, Brain Informatics, pages

558–567, Cham. Springer International Publishing.

Lal, S. K. and Craig, A. (2001). A critical review of the

psychophysiology of driver fatigue. Biological Psy-

chology, 55(3):173–194.

Lal, S. K. and Craig, A. (2002). Driver fatigue: Electroen-

cephalography and psychological assessment. Psy-

chophysiology, 39(3):313–321.

May, J. F. and Baldwin, C. L. (2009). Driver fatigue:

The importance of identifying causal factors of fatigue

when considering detection and countermeasure tech-

nologies. Transportation Research Part F: Trafﬁc Psy-

chology and Behaviour, 12(3):218–224.

Minhas, R., Arbatli, S., Celik, Y., Peker, Y., and Semiz,

B. (2024a). A novel approach to quantify microsleep

in drivers with obstructive sleep apnea by concurrent

analysis of eeg patterns and driving attributes. IEEE

Journal of Biomedical and Health Informatics.

Minhas, R., Peker, N. Y., Hakkoz, M. A., Arbatli, S., Celik,

Y., Erdem, C. E., Semiz, B., and Peker, Y. (2024b).

Association of visual-based signals with electroen-

cephalography patterns in enhancing the drowsiness

detection in drivers with obstructive sleep apnea. Sen-

sors, 24(8).

emcov

a, A., Svozilov

a, V., Bucsuh

azy, K., Sm

sek, R.,

ezl, M., Hesko, B., Bel

ak, M., Bil

ık, M., Maxera,

P., Seitl, M., Dominik, T., Semela, M.,

Sucha, M.,

and Kol

r, R. (2021). Multimodal features for de-

tection of driver stress and fatigue: Review. IEEE

Transactions on Intelligent Transportation Systems,

22(6):3214–3233.

Parui, S., Bajiya, A. K. R., Samanta, D., and Chakravorty,

N. (2019). Emotion recognition from eeg signal using

xgboost algorithm. In 2019 IEEE 16th India Coun-

Electroencephalography Analysis Frameworks for the Driver Fatigue Problem: A Benchmarking Study

835

cil International Conference (INDICON), pages 1–4.

IEEE.

Philip, P., Sagaspe, P., Moore, N., Taillard, J., Charles, A.,

Guilleminault, C., and Bioulac, B. (2005). Fatigue,

sleep restriction and driving performance. Accident

Analysis & Prevention, 37(3):473–478.

Simon, M., Schmidt, E. A., Kincses, W. E., Fritzsche, M.,

Bruns, A., Aufmuth, C., Bogdan, M., Rosenstiel, W.,

and Schrauf, M. (2011). Eeg alpha spindle measures

as indicators of driver fatigue under real trafﬁc condi-

tions. Clinical Neurophysiology, 122(6):1168–1178.

Singh, A. K. and Krishnan, S. (2023). Trends in eeg signal

feature extraction applications. Frontiers in Artiﬁcial

Intelligence, 5.

Wang, F., Wang, H., Zhou, X., and Fu, R. (2022). A driving

fatigue feature detection method based on multifractal

theory. IEEE Sensors Journal, 22(19):19046–19059.

Wei, C.-S., Wang, Y.-T., Lin, C.-T., and Jung, T.-P. (2018).

Toward drowsiness detection using non-hair-bearing

eeg-based brain-computer interfaces. IEEE Transac-

tions on Neural Systems and Rehabilitation Engineer-

ing, 26(2):400–406.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

836