Mental Workload Estimation using Wireless EEG Signals

Quadri Adewale

and George Panoutsos

Montreal Neurological Institute, McGill University, Montreal, Canada

Automatic Control & Systems Engineering, University of Sheffield, Sheffield, U.K.

Keywords: Electroencephalogram (EEG), Mental Workload, Cross-task, Cross-subject, Cross-session, Wireless EEG

Headset, Domain Adaptation, N-Back Task, Mental Arithmetic Task.

Abstract: Previous studies have demonstrated the applicability of electroencephalogram (EEG) in estimating mental

workload. However, developing reliable models for cross-task, cross-subject and cross-session classifications

of workload remains a challenge. In this study, we used a wireless Emotiv EPOC headset to evaluate workload

in eight subjects and two mental tasks, namely n-back, and arithmetic tasks. 0-back and 2-back tasks, and 1-

digit and 3-digit additions were employed as low and high workloads in the n-back and arithmetic tasks,

respectively. Using power spectral density as features, a signal processing and feature extraction framework

was developed to classify workload levels. Within-session accuracies of 98.5% and 95.5% were achieved in

the n-back and arithmetic tasks, respectively. To facilitate real-time estimation of workload, a fast domain

adaptation technique was applied to achieve a cross-task accuracy of 68.6%. Similarly, we obtained accuracies

of 80.5% and 76.6% across sessions, and 74.4% and 64.1% across subjects, in n-back and arithmetic tasks,

respectively. Although the number of participants is limited, this framework generalised well across subjects

and tasks, and provides a promising approach towards developing subject and task-independent models. It

also shows the feasibility of using a consumer-level wireless EEG headset in cognitive monitoring for real-

time estimation of workload in practice.

1 INTRODUCTION

Brain-computer interface (BCI) is mainly applied to

aid disabled persons by using the brain signals for

communication and control while bypassing auxiliary

muscles or nerves (Wolpaw et al., 2002). However,

BCI is now used in healthy subjects in an application

called Passive BCI (Zander et al., 2010). Passive BCI

can be used to obtain information about a user’s level

of workload, mental state, or attentiveness. This

application can help to improve a vehicle driver’s

performance, prevent accidents in systems and

industries and ensure attentiveness of security

officers in surveillance systems (Mueller et al., 2008;

Venthur et al., 2010; Welke et al., 2009).

Although there is no universal definition of

mental workload (Cain, 2007; Zander et al., 2010),

workload can be viewed as the result of the

interaction between work demands and human

capacity (Hart and Staveland, 1988). As the workload

increases, the task demand approaches the upper limit

https://orcid.org/0000-0002-7395-8418

of human ability. Physiological correlates of

workload have been established in many literatures.

Some of these measures include heart rate (Brookings

et al., 1996), blood pressure (Theorell et al., 1988),

Electromyogram (EMG) (Mehta & Agnew, 2012)

and EEG. Even though there seem not to be a best

physiological indicator of workload, some studies

showed EEG to be more promising compared to other

indicators (Hogervorst et al., 2014; Taylor et al.,

2010).

In mental workload classification, a variety of

machine learning methods have been employed.

Some of these methods include support vector

machine (SVM) (Wang et al., 2016), artificial neural

network (ANN) (Baldwin and Penaranda, 2012) and

hierarchical Bayes model (Wang et al., 2012). SVM

finds more application because it generalizes well and

handles high-dimensional data (Burges, 1998).

However, due to the nonstationary nature of EEG

signals, the performance of algorithms degrades when

the training and test data are taken from different

sessions and subjects. Hence such algorithms need to

200

Adewale, Q. and Panoutsos, G.

Mental Workload Estimation using Wireless EEG Signals.

DOI: 10.5220/0010251302000207

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 200-207

ISBN: 978-989-758-490-9

be trained or adapted for every user and session

(personalised and bespoke models). While such data

modelling attempts are useful, not being able to use

previously elicited models in applications for other

users is a weakness in terms of developing multi-user

software tools and algorithms.

Some attempts have been made to overcome this

performance degradation. EEG source localization

and functional connectivity estimation ware applied

to classify workload across tasks (Dimitrakopoulos et

al., 2017). Similarly, a combination of deep recurrent

network and 3D convolutional neural network was

used to learn detailed features for cross-task

classification (Zhang et al., 2019). Other studies

proposed domain adaptation and transfer learning to

overcome the shifts in data distribution across

different subjects (Albuquerque et al., 2019; J. Zhang

et al., 2017). However, these studies considered either

cross-task or cross-subject classifications separately.

Moreover, some of them used many electrodes for

recording the EEG signals which reduces the

comfortability of using EEG headsets in practical and

online scenarios.

We address these issues by (i) applying a single

framework to overcome cross-session, cross-subject

and cross-task performance degradations (ii) using a

consumer-level wireless EEG headset with just 14

channels. We developed a simple signal processing

and feature extraction technique to facilitate practical

and real-time application. The model was tested

across eight (8) subjects in two different types of task

– n-back task and arithmetic task. We then applied a

fast domain adaptation paradigm called Adaptive

Subspace Feature Machine (ASFM) (Chai et al.,

2017) to improve the model performance across

sessions and tasks. We compared the results from

ASFM with those of SVM. Some subjective and

performance indices of mental workload were also

used to verify that the experimental design reflects

different levels of workload.

2 METHOD

2.1 Subjects

Eight (8) subjects (6 males and 2 females)

participated in the EEG experiment which was held at

the Physiological Signal Processing Laboratory,

Department of Automatic Control and Systems

Engineering, University of Sheffield. Participants

were aged between 19 and 30 (Mean = 25 ± 3 years).

All subjects were right-handed, reported normal or

corrected-to-normal vision, and had no history of any

fatigue-related disorder. The experiment was

performed in accordance with the University’s ethics

guidelines, and participants gave written informed

consent.

2.2 Experimental Design

N-back task and arithmetic task were employed in this

study and each task had two difficulty levels. The two

tasks have been extensively used to induce workload

demands (Dimitrakopoulos et al., 2017; S. Wang et

al., 2016; Zarjam et al., 2012). In the n-back task, 0-

back and 2-back tasks were used to represent low and

high workloads, respectively. As shown in Figure 1a,

for the 0-back condition, the target letter is ‘X’. For

the 2-back condition, participant decides if the letter

displayed currently is same as the letter displayed two

sequences earlier. Hence, the participant updates his

memory by memorizing two previous letters as the

sequence progresses. In both task levels, the

participant presses the appropriate key to indicate if

the letter is a target or not.

In the arithmetic tasks, participants are required to

perform arithmetic operations without any aid such as

pen and paper or calculator. The answer from every

arithmetic operation is memorized and retrieved after

some seconds when an answer is displayed. If the

number displayed on the screen is the correct result

from the last arithmetic operation, then such number

is the target number (T), else it is a non-target (NT). 1-

digit addition was used for low workload level and 3-

digit addition for high workload level. The 3-digit

addition is shown in Figure 1b.

To perform the experiment, the participants’

attention is focused on a cross on the screen for 30

seconds without any movement or much eye blinking.

Then, participants perform 5-minute blocks each of

the 0-back task, 2-back, 1-digit arithmetic and 3-digit

arithmetic tasks. To remove time-dependent

confounding effects, 3 participants were asked to

repeat the experiment after a week, while the tasks

were presented in a counterbalanced order. The

participant took a break after every task block and

rated each task on an RSME scale (Zijlstra, 1993)

based on the perceived expended effort in solving that

task. The scale ranges from 0 to 150 in increasing

order of perceived effort expended.

While performing the tasks, the EEG signals of

participants were recorded using a wireless Emotiv

EPOC neuroheadset (EMOTIV, 2013). The Emotiv

EPOC headset uses 14 electrodes with two additional

electrodes for referencing (DRL) and noise

Mental Workload Estimation using Wireless EEG Signals

201

cancellation (CMS). All the available electrodes were

used in this study.

(a)

(b)

Figure 1: Experimental tasks used to evaluate workload

levels. (a) N-back tasks. (b) 3-digit arithmetic task.

2.3 Data Analysis

2.3.1 Data Acquisition and Pre-processing

The EEG data were sampled at 128Hz. All the 14

channels were used, with the reference electrode

attached to the left mastoid. Each raw EEG

measurement was imported into MATLAB and data

corresponding to the fourteen channels were

extracted. A bandpass filter of 1.5-40Hz was applied

to remove high frequency noise and low frequency

DC components. The bandpass filtering was done in

in two directions to avoid phase shift or distortion of

the EEG data. With the aid of the markers set during

the EEG recording, the epochs corresponding to each

task were extracted.

2.3.2 Feature Extraction

The filtered data were divided into 4-second blocks

with 2-second overlaps between adjacent blocks. The

data in each block was normalised for zero mean as

shown in (1) below.

𝒙



𝒙𝒙



(1)

where x is the whole data in a 4-second block, 𝒙



is the mean of the data in such block and 𝒙



the normalised data in the block. The power spectral

density (PSD) in each normalised block was

computed using Welch’s method with 1-second

Hamming window and 50% overlap. Windowing was

necessary to reduce signal leakage, and the overlaps

allow for smooth transition between windows. In

each block, the power spectral densities of eight

frequency bands were computed thus: 4-8Hz, 8-

12Hz, 12-16Hz, 16-20Hz, 20-24Hz, 24-28Hz, 28-

32Hz, 32-36Hz. For each frequency band, the root-

mean-square (RMS) value was calculated as follows:

𝑅𝑀𝑆 

‖

𝑃𝑆𝐷

‖

√

𝑙



)

where ||PSD|| is the Euclidean length of the PSD in a

frequency band and 𝑙



is the length of the PSD vector.

With 14 channels and 8 frequency bands, 112 (14 ×

8) features were generated.

2.3.3 Data Classification

An SVM (with a linear or RBF kernel) was used for

classifying the workload levels in the n-back and

arithmetic tasks. In addition, the performance of the

SVM was investigated for cross-session, cross-task,

and cross-subject classifications. ASFM was also

applied and the results were compared with those

obtained from SVM.

ASFM was proposed in (Chai et al., 2017) as a fast

domain adaptation technique for EEG-based emotion

recognition to overcome the performance degradation

when EEG data are sampled from different subjects

or sessions. The nonstationary nature of EEG and

variability of brain dynamics with individuals and age

cause a mismatch between the marginal and

conditional distributions of the source domain

(training data) and target domain (testing data). In

other words, if there is a source domain 𝑋



with label

and a target domain 𝑋



with label 𝑌



, ASFM

formulates a new feature to reduce the marginal

distribution mismatch between 𝑃



𝑋



and 𝑃



𝑋



, and

conditional distribution mismatch between 𝑃



𝑌



|𝑋





and 𝑃



𝑌



|𝑋



.

First, a Subspace Alignment (Fernando et al.,

2013) is used to unify the marginal distribution of the

source and target domains through principal

component analysis (PCA). The eigenvectors form

the new subspaces 𝑍



and 𝑍



for the source and

target, respectively. A linear transformation is then

obtained to map 𝑍



to 𝑍



. If there is a transformation

B, an objective function can be formulated to align

the subspaces as follow:

𝑚𝑖𝑛 

‖

𝑍



𝐵𝑍



‖







)

where

‖





is Frobenius norm. The above equation

can be rewritten as:







‖

𝑍





𝑍



𝐵𝑍





𝑍



‖







‖

𝐵𝑍





𝑍



‖

)

where T denotes transpose. Hence, the objective

function is minimised when 𝐵

∗

𝑍





𝑍



. Then, the

subspace can be transformed thus:

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

202

𝑍



 𝑍



𝑍





𝑍



(5)

To reduce the marginal distribution mismatch

between 𝑃



𝑋



 and 𝑃



𝑋



 , 𝑋



𝑍



 𝑋



𝑍



𝑍





𝑍



and 𝑋



𝑍



are then computed.

Next, the conditional distributions in 𝑋



𝑍



and

𝑋



𝑍



are adapted by obtaining the probability for an

input target in the transformed subspace and moving

the target to the training set. In other words, the

discrepancy between 𝑃



𝑌



|𝑋



𝑍



 and

𝑃



𝑌



|𝑋



𝑍



 is reduced. If 𝑋



𝑍



is denoted as L,

then the transformed source domain can be

represented as



𝑙



,𝑙



… 𝑙





with labels



𝑦



,𝑦



… 𝑦





Logistic regression is applied to compute the

conditional distribution of the source. Probabilistic

model of logistic regression is given as:

𝑃



𝑦

𝑙;𝑤





1exp



𝑦𝑤



𝑙



(6)

where 𝑦 1 and 𝑤 is a weight vector that can be

learnt with gradient descent algorithm. The

conditional distribution of the source domain can then

be written according to Equation (2.34) as follows:

𝑃





𝑌



1

𝐿;𝑤



 𝑃





𝑌



1

𝑋



𝑍



;𝑤





1exp𝑤



𝑋



𝑍





(7)

On the other hand, the conditional distribution

𝑃



𝑌



|𝑋



𝑍



 of the target domain cannot be exactly

calculated since the target data are not labelled. A

solution is assumed that 𝑃





𝑌



𝑋



𝑍







𝑃





𝑌



𝑋



𝑍



;𝑤



. In an iterative manner, logistic

regression could be used to estimate 𝑃





𝑌

𝑋



𝑍



;𝑤



to indicate the level of certainty that 𝑋



𝑍



belongs to

the predicted label y. Consequently, a confidence

level c is defined to determine the target samples that

would be moved to the training set.

𝑐



𝑍







1 𝑖𝑓 𝑃





𝑌

𝑋



𝑍



;𝑤



 𝜏

0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

(8)

where 𝜏 is a threshold between 0 and 1. Samples with

confidence level of 1 in the target set are moved to the

training set and the conditional distribution

𝑃





𝑌



𝑋



𝑍



;𝑤



is recomputed. As the process is

repeated in more iterations, the marginal distribution

discrepancy between the source and target domain is

reduced.

3 RESULTS

3.1 Subjective Measure (RSME)

The mental workload perceived by subjects increased

with memory load, with average rating of 44 for 0-

back task and 86 for 2-back task on the RSME scale.

Paired-samples t-test showed that the two task levels

were significantly different (t(7) = -9.361, p<0.05).

The average ratings for the 1-digit and 3-digit

arithmetic tasks were 42 and 73 respectively, and the

two task levels differed significantly (t(7) = -4.47,

p<0.05). The averages of both the n-back tasks and

arithmetic tasks confirmed that our experimental

design provides two discriminative levels of

workload (low and high).

3.2 Performance Measures

Average response time increased with workload from

0-back 547.9ms (0-back) to 853.9ms (2-back). Due to

non-normality, we used Wilcoxon signed-rank test

and found a significant difference in response times

of the two workload levels (p = 0.012; p < 0.05).

Response time on the arithmetic task also increased

with workload level from 972.5ms (1-digit) to

1251.8ms (3-digit). The difference was statistically

significant (t(7) = -4.773, p = 0.002; p < 0.05),

implying a significant interaction between the speed

of performance of a task and workload.

The average accuracy of response to stimuli

significantly degraded (t(7) = 3.399, p = 0.011; p <

0.05) as the workload increased from 0-back (98.2%)

to 2-back (91.4%). Increase in workload from 1-digit

to 3-digit arithmetic also resulted in significant

decrease in average accuracy from 92.5% to 78.5% (p

= 0.048; p < 0.05). The results from both tasks

confirmed the expected difference between the

difficulty levels of low and high mental workloads.

3.3 Variation of EEG Spectral Power

Grand averages of spectral powers across all the eight

subjects for some brain regions are shown in Figure

2. In consonance with previous studies, alpha power

(8-12Hz) decreased with workload across all the

electrodes, gamma power (>25Hz) increased with

workload, and theta power (4-7Hz) increased with

workload. Similar to the findings in (S. Wang et al.,

2016), increase in power with workload was observed

in the high beta band (20-25Hz), especially at the

frontal sites (AF4 and FC6). Furthermore, the effect

of workload on spectral power is prominent in the

Mental Workload Estimation using Wireless EEG Signals

203

gamma band across all electrodes. The results support

the use of EEG spectral power as a feature for

estimating mental workload.

(a) (b)

Figure 2: Grand averages of spectral power vary with

workload across frequency bands. (a) Power spectra in

frontal region. (b) Power spectra in temporal region.

3.4 Classification Accuracies

3.4.1 Within-session Classification

The EEG obtained from a subject in an experiment

session was used to train and test the accuracy of the

model for such subject in a 10-fold cross-validation.

Figure 3a shows the performance of the SVM (with

linear kernel) in classifying the workload levels for

the two types of task. The algorithm classified the two

levels of workload in n-back task with an average

accuracy of 98.5% (SD = 2.1%) as against a mean

accuracy of 95.5% (SD = 4.1%) for the two workload

levels in arithmetic task. The average accuracies are

close to the 98.6% (0-back vs 2-back) and 94.2% (1-

digit vs 2-digit multiplication) obtained in (Hwang et

al., 2014) using same Emotiv EPOC headset. About

100% accuracy was also reported by (Wang et al.,

2016) using Emotiv headset for 0-back vs 2-back

tasks.

The highest and lowest accuracies achieved for

the n-back task were 100% and 93.5%, respectively.

The arithmetic task produced 100% and 88.4% as the

highest and lowest accuracies, respectively. The

classification accuracies in the n-back and arithmetic

tasks were significantly different p=0.028 (p<0.05).

The difference in accuracy for the two tasks could be

because be the two levels of workload in the

arithmetic tasks have more similarities than those of

the n-back tasks. Hence, there are likely more

common features in the 1-digit and 3-digit subtasks

which makes it less easy for the algorithm to

discriminate between the two arithmetic workload

levels. It could also be that there are more cross-

subject variabilities in the arithmetic task than the n-

back task, therefore, the model could generalise better

for the latter task. As shown in the results, accuracies

of the two tasks varied between subjects. Very high

accuracies were achieved for subjects P05 and P08.

These discrepancies point to the variation of brain

dynamics with individuals and age; hence, a model

may not generalise well across subjects and therefore

require tuning for every user. However, the model

developed in this work generalised across many

subjects without individual-based tuning.

3.4.2 Cross-session Classification

Due to the nonstationary nature of EEG signals, the

performance of a model degrades if the training and

test data are from different sessions or times. As a

result, training is often repeated for every session. To

test the performance of the model across different

training sessions, three participants were asked to

repeat the tasks after seven days. Then, the data from

the first day were used for training while the data from

the eighth day were used for testing. Here, SVM and

ASFM were applied for classification and compared

against each other as shown in Figure 3b.

The performance of SVM degraded when the

trainings from the previous experiment session were

used to classify data obtained many days later without

retraining. The accuracy of SVM, without any feature

adaptation, reduced to as low as 43.9% (below 50%)

in one of the cases. Conversely, the use of ASFM, a

domain adaption technique, achieved high average

cross-session accuracies of 76.6% (SD = 2.5%) and

80.5% (SD=16%) in the arithmetic and n-back tasks,

respectively.

ASFM reduced the marginal and conditional

distribution mismatch of EEG data across two

different experimental sessions. This result suggests

that the model with ASFM could be used for a subject

at every session without retraining. ASFM was first

applied for emotion recognition using differential

entropy as features (Chai et al., 2017). In that work, it

achieved a cross-session accuracy of 75.1% (SD =

7.7%). Our work has however shown that it can be

successfully applied to mental workload using power

spectral density as features.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

204

(a) (b)

Figure 3: Classification accuracies of the model (a) Within-session classification accuracy using SVM with linear kernel. (b)

Cross-session performance on arithmetic task. (c) Cross-subject performance on arithmetic task. (d) Cross-task classification

accuracies.

Table 1: Average classification accuracies.

3.4.3 Cross-subject Classification

To evaluate the effect of variability of brain dynamics

across subjects, the model was evaluated for cross-

subject performance. Leave-one-subject-out

classification method was applied by using data from

one subject for testing and the data from the

remaining seven subjects for training. The procedure

was repeated eight times so that data from every

subject was used for testing. To limit the size of the

training data, only about 60-second data window (60

samples) was selected from each subject for inclusion

in the training set. Hence, the training set contained

420 samples. In the test set, the whole 5-minute length

of data from a subject was used. Furthermore, the

kernel of the SVM was changed to RBF kernel

because the linear kernel could not find a linear

hyperplane for one of the cases. SVM with RBF

kernel was compared against ASFM as shown in

Figure 3c. SVM achieved a mean classification

accuracy of 60.4% (SD = 20.5%) and 52.6% (SD =

4.2%) on n-back and arithmetic tasks, respectively.

ASFM improved the cross-subject accuracies to

74.4% (SD =13%) and 64.1% (SD = 9.5%) in the n-

back and arithmetic tasks, respectively.

Even though using a non-linear kernel can

improve performance of SVM or even find a solution

where using linear kernel is infeasible, SVM without

feature adaptation is limited in capturing the cross-

human variability that exists in brain dynamics. Such

limitation is observed in subject P01 where the model

performance deteriorated below the average level.

The results show that feature adaptation with ASFM

can mitigate the effect of subject variability on model

performance.

3.4.4 Cross-task Classification

Cross-task performance of the model was examined

by training on n-back tasks and classifying on

arithmetic tasks. The result of the cross-task

classification is shown in Figure 3d. SVM with RBF

kernel provided an average accuracy of 52% (SD =

5.5%) while ASFM yielded a higher average

Cross-Task Acc. (%)

N-Back Arithmetic N-Back Arithmetic N-Back Arithmetic

SVM 98.5 (SD = 2.1) 95.5 (SD = 4.1) 63 (SD = 18.5) 57.6 (SD = 6.2) 60.4 (SD = 20.5) 52.6 (SD = 4.2) 52 (SD = 5.5)

ASFM --- --- 80.5 (SD = 16) 76.6 (SD = 2.5) 74.4 (SD = 13) 64.1 (SD = 9.5) 68.6 (SD =15.8)

Within-Session Acc. (%) Cross-Session Acc. (%) Cross-Subject Acc. (%)

Mental Workload Estimation using Wireless EEG Signals

205

accuracy of 68.6% (SD = 15.8%). The deterioration

in performance could be attributed to the difference

in absolute workload levels in the two tasks. For

example, low workload level in the n-back task (0-

back) may not be equivalent to low workload level in

the arithmetic task (1-digit). This effect is also

observable in the differences of average subjective

ratings on the RSME scale presented earlier. Besides,

the underlying brain dynamics resulting from

performing the n-back tasks could be different from

those of the arithmetic tasks. Nevertheless, the use of

ASFM as a feature adaptation technique reduced the

mismatch between the different workload types. The

classification results are summarised in Table 1.

4 CONCLUSION

This work proposed a robust modelling technique for

online estimation of mental workload using a 14-

channel wireless EEG headset. The subjective and

performance measures indicated that the

experimental design provided discriminative

workload levels. Using SVM with linear kernel, the

model could classify workload levels in more than

one type of task without requiring subject or task

adaptation. Furthermore, a domain adaptation

technique, ASFM, was used to overcome the

variabilities that exist across subjects, experimental

sessions, and tasks. ASFM showed better

performance than SVM (with RBF kernel) in the

presence of these variabilities. ASFM – to the best of

our knowledge – has not been used in estimating

workload before. However, it was successfully

applied in this work and yielded good performance in

cross-subject, cross-session and cross-task

classifications of workload. This research provided a

promising framework for estimating mental workload

across subjects, sessions, and tasks. It also shows the

feasibility of developing models that would not

require retraining or recalibration when there are

changes in users, sessions, or types of task.

In this preliminary study, only 8 subjects were

included in the trials for performance evaluation, and

3 subjects for cross-session classification. Based on

the very promising results obtained, it is

recommended that a larger study is conducted with

more participants to establish the generalisation and

robustness of the proposed method.

In addition, this work has used two separate types

of task to estimate workload, more tasks can be

designed to further investigate the generalisability of

the model across different tasks. In addition, multi-

class workload levels can be used instead of the two-

class workload levels to capture more levels of

workload such as ‘very low’, ‘very high’, etc.

Validation on many tasks and workload levels can

facilitate the development of a task-independent

model for within-task and cross-task classification in

practical settings. The model can also be tested in

real-time when the subjects are performing cognitive

tasks. Ultimately, this research work highlights the

potential for the creation of a robust online cognitive

monitoring system for assessing mental workload in

practical situations.

REFERENCES

Albuquerque, I., Monteiro, J., Rosanne, O., Tiwari, A.,

Gagnon, J. F., & Falk, T. H. (2019). Cross-subject

statistical shift estimation for generalized

electroencephalography-based mental workload

assessment. Conference Proceedings - IEEE

International Conf. on Systems, Man and Cybernetics.

https://doi.org/10.1109/SMC.2019.8914469

Baldwin, C. L., & Penaranda, B. N. (2012). Adaptive

training using an artificial neural network and EEG

metrics for within- and cross-task workload

classification. NeuroImage, 59(1), 48–56. https://

doi.org/10.1016/j.neuroimage.2011.07.047

Brookings, J. B., Wilson, G. F., & Swain, C. R. (1996).

Psychophysiological responses to changes in workload

during simulated air traffic control. Biological

Psychology, 42(3), 361–377.

https://doi.org/10.1016/0301-0511(95)05167-8

Burges, C. J. C. (1998). A Tutorial on Support Vector

Machines for Pattern Recognition. Data Mining and

Knowledge Discovery, 2(2), 121–167. https://

doi.org/10.1023/A:1009715923555

Cain, B. (2007). A Review of the Mental Workload

Literature. Defence Research and Development

Toronto (Canada), 1998, 4-1-4–34. http://

www.dtic.mil/cgi-bin/GetTRDoc?Location=U2&do c

=GetTRDoc.pdf&AD=ADA474193

Chai, X., Wang, Q., Zhao, Y., Li, Y., Liu, D., Liu, X., &

Bai, O. (2017). A fast, efficient domain adaptation

technique for cross-domain electroencephalography

(EEG)-based emotion recognition. Sensors

(Switzerland), 17(5), 1–21. https://doi.org/10.339

0/s17051014

Dimitrakopoulos, G., Kakkos, I., Dai, Z., Lim, J.,

Bezerianos, A., & Sun, Y. (2017). Task-independent

mental workload classification based upon common

multiband EEG cortical connectivity. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, 4320(c), under revision. https://doi.

org/10.1109/TNSRE.2017.2701002

EMOTIV. (2013). EMOTIV Epoc - 14 Channel Wireless

EEG Headset. EMOTIV Inc.

Fernando, B., Habrard, A., Sebban, M., & Tuytelaars, T.

(2013). Unsupervised visual domain adaptation using

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

206

subspace alignment. Proceedings of the IEEE

International Conference on Computer Vision, 2960–

2967. https://doi.org/10.1109/ICCV.2013.368

Hart, S. G., & Staveland, L. E. (1988). Development of

NASA-TLX (Task Load Index): Results of Empirical

and Theoretical Research. Advances in Psychology,

52(C), 139–183. https://doi.org/10.101 6/S0166-

4115(08)62386-9

Hogervorst, M. A., Brouwer, A. M., & van Erp, J. B. F.

(2014). Combining and comparing EEG, peripheral

physiology and eye-related measures for the assessment

of mental workload. Frontiers in Neuroscience,

8(OCT). https://doi.org/10.3389/fnin s.2014.00322

Hwang, T., Kim, M., Hwangbo, M., & Oh, E. (2014).

Comparative analysis of cognitive tasks for modeling

mental workload with electro encephalogram. Conf.

Proceedings : Annual International Conference of the

IEEE Engineering in Medicine and Biology Society.

Annual Conference, 2014, 2661–2665.

https://doi.org/10.1109/EMB C.2014.6944170

Mehta, R. K., & Agnew, M. J. (2012). Influence of mental

workload on muscle endurance, fatigue, and recovery

during intermittent static work. European Journal of

Applied Physiology, 112(8), 2891–2902. https://doi.

org/10.1007/s00421-011-2264-x

Mueller, K. R., Tangermann, M., Dornhege, G., Krauledat,

M., Curio, G., & Blankertz, B. (2008). Machine

learning for real-time single-trial EEG-analysis: From

brain-computer interfacing to mental state monitoring.

Journal of Neuroscience Methods, 167(1), 82–90.

https://doi.org/10.1016/j.jneu meth.2007.09.022

Taylor, G., Reinerman-Jones, L., Cosenzo, K., &

Nicholson, D. (2010). Comparison of Multiple

Physiological Sensors to Classify Operator State in

Adaptive Automation Systems. Proceedings of the

Human Factors and Ergonomics Society Annual

Meeting, 54(3), 195–199. https://doi.org/10.1177

/154193121005400302

Theorell, T., Perski, A., Akerstedt, T., Sigala, F., Ahlberg-

Hultén, G., Svensson, J., & Eneroth, P. (1988). Changes

in job strain in relation to changes in physiological state.

A longitudinal study. Scandinavian Journal of Work,

Environment & Health, 14(3), 189–196.

https://doi.org/10.5 271/sjweh.1932

Venthur, B., Blankertz, B., Gugler, M. F., & Curio, G.

(2010). Novel applications of BCI technology:

Psychophysiological optimization of working

conditions in industry. Conference Proceedings - IEEE

International Conference on Systems, Man and

Cybernetics, 417–421. https://doi.org/10.1109/

ICSMC.2010.5641772

Wang, S., Gwizdka, J., & Chaovalitwongse, W. A. (2016).

Using Wireless EEG Signals to Assess Memory

Workload in the n-Back Task. IEEE Transactions on

Human-Machine Systems, 46(3), 424–435.

https://doi.org/10.1109/THMS.2015.2476818

Wang, Z., Hope, R. M., Wang, Z., Ji, Q., & Gray, W. D.

(2012). Cross-subject workload classification with a

hierarchical Bayes model. NeuroImage, 59(1),64–69

https://doi.org/10.1016/j.neuroimage.2011.07.094

Welke, S., Juergensohn, T., & Roetting, M. (2009). Single-

trial detection of cognitive processes for increasing

traffic safety. Proceedings of the 21st International

Technical Conference on the Enhanced Safety of

Vehicles Conference (ESV)., 1–10.

Wolpaw, J. R., Birbaumer, N., McFarland, D. J.,

Pfurtscheller, G., & Vaughan, T. M. (2002). Brain-

computer interfaces for communication and control.

Clinical Neurophysiology : Official Journal of the

International Federation of Clinical Neurophysiology,

113(6), 767–791. https://doi.org/ 10.1016/S1388-

2457(02)00057-3

Zander, T., Kothe, C., Jatzev, S., & Gaertner, M. (2010).

Enhancing Human-Computer Interaction with Input

from Active and Passive Brain-Computer Interfaces.

Brain-Computer Interfaces, 149–178. https://doi.org

/10.1007/978-1-84996-272-8

Zarjam, P., Epps, J., Lovell, N. H., & Chen, F. (2012).

Characterization of memory load in an arithmetic task

using non-linear analysis of EEG signals. Engineering

in Medicine and Biology Society (EMBC), 2012 Annual

International Conference of the IEEE, 3519–3522.

https://doi.org/10.1109/ EMBC.2012.6346725

Zhang, J., Wang, Y., & Li, S. (2017). Cross-subject mental

workload classification using kernel spectral regression

and transfer learning techniques. Cognition,

Technology and Work. https://doi.org/ 10.1007/s10111-

017-0425-3

Zhang, P., Wang, X., Zhang, W., & Chen, J. (2019).

Learning Spatial-Spectral-Temporal EEG Features

With Recurrent 3D Convolutional Neural Networks for

Cross-Task Mental Workload Assessment. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering. https://doi.org/10.1109/TNSRE.2018

.2884641

Zijlstra, F. R. (1993). Efficiency in work behaviour: A

design approach for modern tools. Delft University

Press, January 1993, 1–186. https://doi.org/90-6275-

918-1

Mental Workload Estimation using Wireless EEG Signals

207