Comfort Assessment Method of EEG-Based Exoskeleton

Walking-Assistive Device

Heyuan Wang

, Kaitai Li

, Hui Liu

, Xuesong Ye

and Congcong Zhou

College of Biomedical Engineering & Instrument Science, Biosensor National Special Laboratory, Zhejiang University,

Hangzhou, Zhejiang, 310027, China

Cognitive Systems Lab, University of Bremen, 28359 Bremen, Germany

Sir Run Run Shaw Hospital, School of Medicine, Zhejiang University, National Engineering Research Center for

Innovation and Application of Minimally Invasive Devices, East Qingchun Road, Hangzhou, Zhejiang, 310016, China

Keywords: Biosignal Processing, Comfort Assessment, EEG, Deep Learning, Exoskeleton Walking-Assistive.

Abstract: The study of wearable exoskeleton robotics has garnered significant attention, amidst a rapidly expanding

corpus of scholarly work aimed at the empirical evaluation of the performance characteristics of robotic

exoskeletons. However, quantifying comfort performance is still a significant and challenging task. This study

aimed to perform comfort assessment based on EEG (Electroencephalography) signals and classical machine

learning models as well as deep learning model. It involved collecting EEG data from users wearing lower

limb exoskeleton walking-assistive devices for comfort assessment during walking experiments. The

subjective evaluation labels of comfort were obtained using a semantic differential scale, providing comfort

labels data for each participant in each trial. This study conducted a comparative analysis of three classical

ML (Machine Learning) models, Naive Bayes, K-Nearest Neighbors, and Support Vector Machine models,

with DL (Deep Learning) model, LSTM (Long Short-Term Memory), in terms of their accuracy for comfort

assessment. The results of the analysis showed that the deep learning model, LSTM, outperformed the

classical machine learning models, in terms of accuracy for evaluating comfort. Specifically, we get an

accuracy of 0.91±0.12 on the LSTM model. The LSTM model demonstrated higher accuracy and better

performance in capturing complex patterns and relationships within the EEG data, leading to the potential of

more accurate predictions of comfort levels.

1 INTRODUCTION

As devices for human-machine symbiosis,

exoskeletons with the strategy of human-in-the-loop,

which incorporates human physiological indicators as

feedback control parameters into the control loop,

have achieved many positive results (Han et al.,

2021a; Song & Collins, 2021; Zhang et al., 2017a). In

the performance evaluation study of lower limb

exoskeleton walking-assistive devices, comfort is

defined as the user's perception of human-robot

interaction (Pinto-Fernandez et al., 2020). Wrong

parameter setting of the assistive device can lead to

discomfort and even pain, which may start fighting

https://orcid.org/0000-0002-1817-9198

https://orcid.org/0000-0002-0680-9818

https://orcid.org/0000-0002-6850-9570

https://orcid.org/0000-0001-8397-1491

the device or engage in other compensatory actions

(Felt et al., 2015). In previous studies on the

optimization of exoskeleton walking-assistive

devices, it has been mentioned that quantifying

comfort is a more challenging direction (Koller et al.,

2016). Powered ankle exoskeleton providing too high

peak torque may lead to discomfort during walking,

and uncomfortable exoskeleton plantarflexion

assistance will increase tibialis anterior muscle

activity (Han et al., 2021b). Some assistive patterns

of the exoskeleton are uncomfortable for the wearers,

resulting in unpleasant optimization experiences and

inaccurate outcomes, so some fixed parameters need

to be set to ensure comfort based on pilot tests (Wang

Wang, H., Li, K., Liu, H., Ye, X. and Zhou, C.

Comfort Assessment Method of EEG-Based Exoskeleton Walking-Assistive Device.

DOI: 10.5220/0012564800003657

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

In Proceedings of the 17th Inter national Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 675-682

ISBN: 978-989-758-688-0; ISSN: 2184-4305

675

et al., 2022). The aim of this paper is to objectively

assess the comfort and later introduce it into the

human-in-the-loop optimization of lower limb

exoskeleton control.

The human body receives various external stimuli

and generates conscious judgments of comfort in the

brain. By assessing brain activity based on

electromagnetic or metabolic activity, it is possible to

evaluate the state of brain activity and objectively

reflect the level of comfort. Commonly used methods

for detecting neural activity include EEG, MEG,

ECoG (Lorach et al., 2023), fNIRS, fMRI (Heeger &

Ress, 2002), etc., the first three are to detect the

electromagnetic activity of the brain, and the last two

are to detect the metabolic activity of the brain. EEG

is one of the most commonly used methods for

measuring brain neural activity. It can detect cluster

electrical activity in the cerebral cortex.

Previous research has shown that changes in

metabolic activity caused by blood circulation can

affect brain electrical rhythms. Additionally, exo-

skeleton assistive devices alter lower limb dynamics

during human walking. This impact can be reflected to

some extent in the comfort level by using several

frequency domain features, especially those related to

the 𝛼 frequency band (Jeong & Kim, 2009;Ling & Xia,

2015;Liu & Chen, 2015;Luo et al., 2020;Park & Lee,

2021). Some researchers have used two classical

machine learning models and EEG signals to evaluate

comfort, and have achieved a classification accuracy of

up to 0.75~0.85 in the binary classification task (Ortiz

et al, 2021). Their study also showed that several

electrodes that were selected to be more relevant to

differentiating comfort when walking were electrodes

located in the primary motor cortex and somatosensory

cortex. This may be related to the difference in gait due

to discomfort. In recent years, deep learning networks

have performed very well on many learning models.

Our research group has conducted study on ankle-foot

motion recognition based on sEMG (surface

Electromyography) and acceleration signals using

classic machine learning models and deep learning

networks (Zhou et al, 2021). The results have

demonstrated the effectiveness of deep learning

networks in processing bioelectric signals. Under the

condition of sufficient data, we have the opportunity to

capture the hidden features that are difficult to be

directly calculated by traditional feature engineering.

In response to the above situation, we conducted

a comfort assessment experiment for exoskeleton

walking-assistive devices based on EEG signals and

classical machine learning models as well as deep

learning model. We validated and compared the

performance of different models in comfort

assessment.

2 METHODOLOGY

This study recruited four healthy university student

participants, from which 10800000 raw data frames

were obtained. The inclusion criteria for recruitment

were as follows: all participants should have no limb

injury, no joint disease, no muscle disease, no nervous

system disease and in good physical condition in the

last week. This series of experiments obtained

approval from the ethics committees of both the

university and the hospital. All participants

volunteered to take part in this study and were

provided with full information about the experimental

setup and procedures before the start of the

experiment.

EEG recording was performed using the

BrainProduct actiCHamp Plus 64-channel device

with a sampling frequency of 2500Hz. The EEG

electrode placement followed the international

standard EEG 10-10 system. The electrodes utilized

actiCAP active electrodes with Ag/AgCl sensors,

providing improved recording capabilities with lower

noise levels. Subjects were instructed to minimize eye

blinking during the experiment and focus on

performing lower limb movements on a treadmill.

Figure 1: Subjects wearing laboratory-developed powered

ankle exoskeleton and unpowered ankle exoskeleton, AFO.

The experimental procedure consists of three parts. In

the first part, the recording of EEG signals begins

while the participants maintain a resting state by

standing still for 15 seconds. The second part involves

the initiation phase of the treadmill, where the

treadmill gradually accelerates to the desired speed

within 15 seconds. The third part involves continuous

walking, which is the phase where vali data is

recorded and lasts for 2 minutes. Three categories of

semantic difference scales (1~2 uncomfortable, 3~5

neutral, and 6~7 comfortable) are set and let the

subjects self-rate after each trial.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

676

Figure 2: EEG preprocessing pipeline: 1)Remove unused EXG channels; 2)1~40Hz filtering; 3)Re-referencing to CAR;

4)FastICA; 5)ICLabel; 6)Remove eye and channel noise ICs.

In the walking experiment, subjects were asked

to wear three different devices: powered ankle

exoskeletons，unpowered ankle exoskeleton (Zhou

et al, 2022), and self-assessed comfortable shoes. The

experiments were conducted at 1.5 m/s, 2.5 m/s, and

3.5 m/s. The powered ankle exoskeleton used in the

experiment was a laboratory-developed device

designed to provide assistance during continuous

walking. Based on ankle joint biomechanics, the

exoskeleton provided a peak assistive torque during

each gait cycle (Cappellini et al., 2006;Zhang et al.,

2017b), while the unpowered ankle exoskeleton was

the Ober AFO.

The subject's EEG signals are recorded by the

BrainProduct EEG device and transmitted to the PC.

Preprocessing of EEG signals is a series of steps that

perform processing on raw EEG data to extract useful

information and remove noise. First, the unused

channel EXG is removed, which is configured to

collect other electrophysiological signals such as

EMG and ECG. Next, filtering in the frequency band

of interest is to highlight the signal in our frequency

range of interest, specifically 1~40Hz bandpass

filtering. In the next step, re-referencing to the CAR

(Common Average Reference) is a commonly used

modified reference method. This method eliminates

common pattern noise between electrodes by

calculating the average of all electrodes and

subtracting the signal from each electrode. This helps

to make the individual electrodes more independent

of each other for better analysis and interpretation of

the EEG signals.

FastICA (Independent Component Analysis)

can be used to extract ICs (Independent Components)

from EEG data, separating the EEG activity that is

mixed together, allowing us to study and understand

the different EEG components. FastICA excels in

both separation quality and computational speed, two

performance metrics that ICA algorithms focus on.

ICLabel is a machine-learning-based tool trained

using a large number of labeled EEG data for

automatically classifying and labeling independent

components in EEG data. It can identify different

types of components in the EEG signal, such as line

noise, channel noise, ocular electrical activity,

electromyographic activity, etc. Finally, based on the

labeling results, the eye and channel noise

components are removed. EEG is filtered to each

frequency band for frequency domain feature

extraction. Filtered EEG band definition: 𝜃(5~7Hz);

𝛼(8~14Hz); 𝛽(15~30Hz); 𝑡𝑜𝑡𝑎𝑙(5~40Hz).

3 FEATURE EXTRACTION

3.1 Frequency Domain Feature

For periodic signals, their Fourier transform

converges, allowing them to be described using a

frequency spectrum. The frequency spectrum of a

periodic signal is distinct and provides an accurate

representation of the signal's components and energy

distribution across different frequencies. The PSD

(Power Spectral Density) is used to describe the

frequency spectrum of actual signals, which are

mostly random signals with infinite energy that

cannot satisfy the absolute integrability condition

required for Fourier transform convergence.

However, PSD cannot be obtained accurately and can

only be estimated using spectral estimation methods.

The PSD of the EEG in specific frequency bands and

the ratio of the PSDs between different frequency

bands are commonly used features of EEG analysis

Comfort Assessment Method of EEG-Based Exoskeleton Walking-Assistive Device

677

(Jap et al., 2009). The AR (Autoregressive) model is

a commonly used parametric method for power

spectrum estimation. The difference equation and

power spectral density of the AR model are as

follows:

𝑥

(

𝑛

)

=−𝑎







𝑥

(

𝑛−𝑘

)

+𝑤

(

𝑛

)

(1)

𝑃



(

𝜔

)

𝜎





1 + 𝑎



𝑒



+⋯+𝑎



𝑒







(2)

Where 𝑤

(

𝑛

)

is a white noise signal with zero

mean and variance of 𝜎





, 𝑝 is the order of the AR

model.

The AR power spectral density estimation method

based on the Burg algorithm was used to extract

feature1 to feature10. The order of the AR model was

set at 18, and the extraction was performed at

intervals of 0.5Hz within the frequency range of 5Hz

to 50Hz. After calculating the PSD for each frequency

band, the following features were obtained:

∑

𝛽/

∑

(𝜃+ 𝛼);

∑

𝛽/

∑

𝛼; max

(

𝛼

)

∑

𝑡𝑜𝑡𝑎𝑙; max

(

𝛽

)

∑

𝑡𝑜𝑡𝑎𝑙;

∑

𝛼/

∑

𝑡𝑜𝑡𝑎𝑙;

∑

𝛽/

∑

𝑡𝑜𝑡𝑎𝑙; max

(

𝜃+𝛼

)

∑

𝑡𝑜𝑡𝑎𝑙 ;

∑

(𝜃+ 𝛼)/

∑

𝑡𝑜𝑡𝑎𝑙 ; max (𝛼) ; max (𝛽) .

The AR power spectral density estimation method

based on the Yule-Walker algorithm, with an AR

model order of 2, was used to extract feature11:

max (𝑡𝑜𝑡𝑎𝑙).

3.2 Time Domain Feature

Standard deviation is a measure of the amount of

variation or dispersion in a set of values. It quantifies

the amount of variation or dispersion of a set of values

from the mean. Standard deviation is calculated for

pre-processed multi-channel EEG data:

𝑠𝑡𝑑=



∑|

𝑥



−𝑥







𝑛−1

(3)

Signal energy is the total power contained in a

signal over a period of time. It is calculated by

integrating the squared values of the signal amplitude

over time and is proportional to the duration and

amplitude of the signal. Calculate Energy from pre-

processed multi-channel EEG data:

𝐸= 𝑥









(4)

The Weibull distribution has two parameters: the

shape parameter and the scale parameter. In this study

rectified EEG were used and only the scale parameter

was calculated. The scale parameter determines the

shape and scale of the distribution. When the scale

parameter increases, the distribution becomes more

concentrated with a higher peak value. Conversely, as

the scale parameter decreases, the distribution

becomes flatter with lower peaks.

𝑓

(

𝑥;𝜆,𝑘

)

=

𝑘

𝜆



𝑥

𝜆





𝑒



(



⁄)



𝑥≥0

0𝑥<0

(5)

where 𝜆 is the scale parameter.

The Hjorth parameter is a set of three time-domain

features that describe a single EEG channel. These

features are activity, mobility, and complexity. They

are commonly used in EEG signals for feature

extraction (Chen et al., 2023;Rizal et al, 2022).

Calculate the Hjorth parameter for the preprocessed

multi-channel EEG data:

𝐻𝑗𝑜



𝑟𝑡ℎ (𝑎𝑐𝑡𝑖𝑣𝑖𝑡𝑦)=𝜎





(6)

𝐻𝑗𝑜



𝑟𝑡ℎ (𝑚𝑜𝑏𝑖𝑙𝑖𝑡𝑦)=𝜎





𝜎



⁄

(7)

𝐻𝑗𝑜



𝑟𝑡ℎ

(

𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦

)

𝜎





𝜎





⁄

𝜎





𝜎



⁄

(8)

Where 𝜎



, 𝜎



, and 𝜎



are the stand deviations

of 𝑥(𝑛), 𝑥′(𝑛), and 𝑥′′(𝑛) respectively. 𝑥(𝑛)

represents the preprocessed sequence of the EEG

signal, and 𝑥′(𝑛) and 𝑥′′(𝑛) represent its first- and

second-order differences.

3.3 Nonlinear Feature

Entropy, originating from Shannon's information

theory as: −

∑

𝑝𝑙𝑜𝑔(𝑝), also known as information

entropy, has given rise to many features in EEG

analysis (Aydın et al., 2009). It is a type of nonlinear

feature. Log energy entropy is a commonly used EEG

feature similar to wavelet entropy, it only involves the

summation of probabilities using logarithms. The

formula is as follows:

𝐻



(

𝑥

)

=−



𝑙𝑜𝑔



𝑝



(

𝑥

)











(9)

Where 𝑝



(𝑥) is probability distribution function

of EEG signal 𝑥, 𝑖 indicates one of the discrete states.

The Sample Entropy (SamEn) is an extension of

the Approximate Entropy (ApEn), which is used to

measure the probability of generating a new pattern

in the signal. The formula is as follows:

𝐻



(

𝑛,𝑚,𝑟

)

=−𝑙𝑛

𝐵



(

𝑟

)

𝐵



(

𝑟

)



(10)

Where 𝑚 is the dimension, which can be 1 or 2; 𝑟

is the approximate tolerance, 𝐵



(

𝑟

)

is the ratio of the

approximate quantity to the total quantity.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

678

4 DATASET AND CLASSIFIERS

4.1 Dataset

4.1.1 Dataset Construction

In order to avoid the influence of the magnitude of

different feature vectors on the results, they were

standardized. After feature extraction and

standardization, the data is shown as Table 1, where

the data frames represent the time frames of the

multichannel EEG, the feature frames represent the

calculated features for each epoch, 5s for each epoch:

Table 1: Exoskeleton walking-assistive device comfort

assessment dataset.

Comfort level

umber of data frames

Uncomfortable 180(2250000)

eutral 540(6750000)

Comfortable 144(1800000)

Feature Frames (Raw Data Frames)

4.1.2 Data Visualization

t-SNE (t-Distributed Stochastic Neighbor

Embedding) is a popular dimensionality reduction

technique used for visualizing high-dimensional data

in a lower-dimensional space. It is particularly

effective in revealing the underlying structure and

patterns within the data (Heggs et al., 2023).

Figure 3: Schematic diagram of t-SNE dimensionality

reduction of the three classifiers, from top to bottom, are

Naive Bayes, KNN and SVM classifiers, in which red

represents ‘uncomfortable’, green represents ‘neutral’, and

blue represents ‘comfortable’.

The algorithm works by constructing a probability

distribution over pairs of high-dimensional data

points, both in the original space and in the lower-

dimensional space. It then tries to minimize the

divergence between these two distributions. In

simpler terms, t-SNE aims to find a lower-

dimensional representation that maintains the

similarities between data points from the original

high-dimensional space, while also ensuring that

dissimilar points are well-separated. By iteratively

optimizing this objective function, t-SNE gradually

maps the data points into the lower-dimensional

space, where they can be visualized and analysed

effectively. In the study of decoding handwritten

characters through an intracortical brain-computer-

interface, 31 handwritten characters can be clearly

distinguished by the two-dimensional visualization of

neural activity drawn by t-SNE (Willett et al., 2021).

As can be seen from the visualization results in

Figure 3, several classifiers have a certain degree of

distinction, and the distinction degree of KNN and

SVM is obviously higher than that of the Naive Bayes

model, the intra-class distance is smaller, and the

inter-class distance is larger.

4.2 Classifiers

4.2.1 Naive Bayes

Naive Bayes model is an elementary yet efficient

algorithm (Wickramasinghe & Kalutarage, 2021) that

assumes independence among features and leverages

prior probabilities for sample class determination. It

is widely used in text classification, sentiment

analysis, and more due to its simplicity and

effectiveness.

4.2.2 SVM

SVM model using ECOC (Error-Correcting Output

Code) can help with multi-class classification

problems (Ubeyli, 2008), encoding categories into

binary codes and building binary SVM classifiers for

each code. The final classification result is

determined by the output encoding of these binary

classifiers.

4.2.3 KNN

KNN is an instance-based model that classifies a new

sample based on its nearest neighbor's category,

calculated using the predefined number of neighbors

K. KNN is simple and performs well on small

datasets, but may face computational and storage

challenges with large datasets (Bablania et al., 2018).

Comfort Assessment Method of EEG-Based Exoskeleton Walking-Assistive Device

679

4.2.4 LSTM

LSTM is a type of RNN (Recurrent Neural Network)

used for processing and predicting time series data. It

has stronger memory capabilities and handles long-

term dependencies better than traditional RNNs.

LSTM uses gates to control information flow,

including forget, input, and output gates. These gates

update cell states, enabling the network to learn long-

term dependencies effectively and avoid gradient

problems during training. In past EEG studies, LSTM

has shown good performance (Du et al., 2022). In this

study, time series data before feature extraction is

used in the LSTM model.

5 RESULTS

5.1 Model Optimization

The K value in the KNN model indicates how many

data points of the nearest neighbors are considered

when classifying. Specifically, when a new data point

needs to be classified or predicted, the KNN

algorithm finds the data points in the K training set

that are closest to that data point, and then uses them

to determine the classification or prediction value of

the new data point. The choice of K value has a great

impact on the performance and results of the KNN

model. To balance the number of frames between the

different classes, stratified random sampling is used

to divide the training set (60%) and the test set (40%).

The test results show that the k value of the optimal

KNN classifier ranges from 5 to 15.

Figure 4: The change curves of classification accuracy,

precision, recall and F1 score of the KNN model trained

after a stratified random sampling of the data set with the

increase of k value.

Figure 4 compares the change curves of various

multiclassification evaluation metrics with the

change of K-value, including accuracy, macro-

average precision, macro-average recall, macro-

average F1 score, weighted-average precision, and

weighted-average F1 score, which are all possible

metrics. The optimal k value of 5 of the 6 evaluation

metrics is 7. In this stratified random sampling, the

optimal accuracy of the model is 84.9%. After 20

times of stratified random sampling, grid search was

used to find the optimal classifier at k=5~15, and the

average optimal accuracy was 88.32%±1.2%.

5.2 Comparison of Classical ML

Models

In a stratified random sampling, the performance of

the three models is calculated and compared. The

performance of SVM and KNN model is significantly

higher than that of Naive Bayes model. The

performance of SVM and KNN model is similar, but

in general, SVM model is slightly better than KNN

model. Specific comparisons are shown in Table 2,

Table 3 and Table 4 below.

Table 2: SVM model performance.

Class

Metric Accuracy Precision Recall F1 Score

Uncomfor

able

0.9267 0.8235 0.718 0.7671

Neutral

0.8879 0.8978 0.9111 0.9044

Comfor

able

0.9526 0.8853 0.931 0.9076

SVM model accuracy: 88.36%

Table 3: KNN model performance.

Class

Metric Accuracy Precision Recall F1 Score

Uncomfor

able

0.9353 0.8 0.8205 0.8101

Neutral

0.8664 0.8662 0.9111 0.8881

Comfor

able

0.931 0.92 0.7931 0.8519

KNN model accuracy: 86.64%

Table 4: Naive Bayes model performance.

Class

Metric Accuracy Precision Recall F1 Score

Uncomfor

able

0.8621 0.6667 0.359 0.4667

Neutral

0.681 0.6784 0.8593 0.7582

Comfor

able

0.7759 0.575 0.3966 0.4694

Naive Bayes model accuracy: 65.95%

It can be seen that although SVM model

outperforms KNN model on most evaluation metrics,

several metrics are lower than KNN model. They are

Uncomfortable Accuracy, Comfortable Precision,

Uncomfortable Recall, Uncomfortable F1 Score.

The accuracy of the Naive Bayes model is not too

low for each class. But Comfortable Recall,

Uncomfortable recall, Comfortable F1 Score,

Uncomfortable F1 Score are all very low.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

680

5.3 ML and DL Models Comparison

Table 5 below compares the performance of the deep

learning model LSTM with three classical machine

learning models. It can be seen that the LSTM model

is superior to the machine learning models in terms of

accuracy of up to 0.91±0.12, even if the optimal KNN

classifier is trained many times, the average accuracy

of these multiple optimal KNN classifiers is only

0.88, and the LSTM can obtain an accuracy of up to

0.91 even without model optimization, indicating that

the LSTM can capture the hidden features of the

relationship between human EEG data and comfort

levels.

Table 5: Comparison of the performance of three classical

Machine Learning models and Deep Learning model

LSTM in this task. The Naive Bayes model was not

optimized. The accuracy of KNN is the average of multiple

optimal KNN models, which is the result in 5.1. The SVM

model with ECOC uses a conventional linear kernel, was

also not optimized. Both Bayes and SVM models were

obtained after 20 stratified random sampling.

Results Models

LSTM Ba

es K

N SVM

Accuracy 0.91±

0.12

0.63±

0.03

0.88±

0.01

0.86±

0.02

(mean±std)

6 CONCLUSIONS

This study is intended to explore the comfort

assessment method for exoskeleton walking-assistive

devices. EEG signals were collected from subjects

during a walking experiment, and frequency-domain,

time-domain, and nonlinear features were extracted.

The t-SNE technique was used for dimensionality

reduction and visualization of categories,

demonstrating separability between different

categories. Subsequently, evaluations based on

several classical machine learning models were

conducted and compared with the performance of the

deep learning model LSTM. The results indicate that

among the classical machine learning models, the

Naive Bayes model performed the worst, with

accuracy far lower than SVM and KNN. Both SVM

and KNN demonstrated good performance, achieving

accuracies above 0.8. The deep learning model LSTM

outperformed several classical machine learning

models in accuracy (0.91±0.12). This indicates that

the deep learning model LSTM exhibits excellent

performance in revealing the potential relationship

between EEG and comfort levels, and can identify

hidden features.

ACKNOWLEDGEMENTS

This work was supported in part by Zhejiang

Provincial Natural Science Foundation of China

under Grant No. LY22H180006 and the National Key

R&D Program of China under Grant No.

2017YFF0210803.

REFERENCES

Aydın, S., Saraoğlu, H. M., & Kara, S. (2009). Log Energy

Entropy-Based EEG Classification with Multilayer

Neural Networks in Seizure. Annals of Biomedical

Engineering, 37, 2626-2630

Bablania, A., Edlaa, D. R., & Dodiaa, S.

(2018Classification of EEG Data using k-Nearest

Neighbor approach for Concealed Information Test.

Paper presented at the 8th International Conference on

Advances in Computing and Communication (ICACC-

2018).

Cappellini, G., Ivanenko, Y. P., Poppele, R. E., &

Lacquaniti, F. (2006). Motor Patterns in Human

Walking and Running. Journal of Neurophysiology,

95(6), 3426-3437

Chen, W., Cai, Y., Li, A., Su, Y., & Jiang, K. (2023). EEG

feature selection method based on maximum

information coefficient and quantum particle swarm.

Scientific Reports, 13.

Du, X., Ma, C., Zhang, G., Li, J., Lai, Y., Zhao, G.,... Wang,

H. (2022). An Efficient LSTM Network for Emotion

Recognition From Multichannel EEG Signals. IEEE

Transactions on Affective Computing, 13(3).

Felt, W., Selinger, J. C., Donelan, J. M., & Remy, C. D.

(2015). "Body-In-The-Loop": Optimizing Device

Parameters Using Measures of Instantaneous Energetic

Cost. PLoS ONE, 10(8)

Han, H., Wang, W., Zhang, F., Li, X., Chen, J., Han, J.,...

Zhang, J. (2021a). Selection of Muscle-Activity-Based

Cost Function in Human-in-the-Loop Optimization of

Multi-Gait Ankle Exoskeleton Assistance. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, 29, 944-952.

Han, H., Wang, W., Zhang, F., Li, X., Chen, J., Han, J.,...

Zhang, J. (2021b). Selection of Muscle-Activity-Based

Cost Function in Human-in-the-Loop Optimization of

Multi-Gait Ankle Exoskeleton Assistance. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, 29, 944-952.

Heeger, D. J., & Ress, D. (2002). What does fMRI tell us

about neuronal activity? Nature Reviews Neuroscience,

3, 142-151

Heggs, R. M., Prado, S., Gava, G. P., Go, M. A., & Schultz,

S. R. (2023). Neural manifold analysis of brain circuit

Comfort Assessment Method of EEG-Based Exoskeleton Walking-Assistive Device

681

dynamics in health and disease. Journal of

Computational Neuroscience, 51, 1-21.

Jap, B. T., Lal, S., Fischer, P., & Bekiaris, E. (2009). Using

EEG spectral components to assess algorithms for

detecting fatigue. Expert Systems with Applications, 36,

2352-2359

Jeong, J., & Kim, H. (2009). Assessment of the Wear

Comfort of Outdoorwear by ECG and EEG Analyses.

Journal of the Korean Society of Clothing and Textiles,

33(10), 1665-1672

Koller, J. R., Gates, D., Ferris, D. P., & Remy, C.

(2016'Body-in-the-Loop' Optimization of Assistive

Robotic Devices: A Validation Study. Paper presented

at the Robotics: Science and Systems XII.

Ling, Y., & Xia, L. (2015). Study on Wearing Pressure

Comfort Based on Females' Physiological Index.

Chinese Journal of Ergonomics, 21(6)

Liu, Y., & Chen, D. (2015). An analysis on EEG power

spectrum under pressure of girdle. International

Journal of Clothing Science and Technology, 27(4),

495-505

Lorach, H., Galvez, A., Spagnolo, V., Martel, F., Karakas,

S., Intering, N.,... Courtine, G. (2023). Walking

naturally after spinal cord injury using a brain–spine

interface. Nature, 618, 126-133.

Luo, X., Yun, S., Xue, C., Yue, Z., & Ren, H. (2020). Study

on comfort evaluation methods of different women's

high-heeled shoes based on electroencephalograph

(EEG) technology. Leather and Footwear Journal,

20(1), 5-14.

Ortiz, M., Vicente, P., Ianez, E., Montiel, E., & Azorin, J.

M. (2021). Assessing Footwear Comfort by

Electroencephalography Analysis. IEEE Access, 9,

134259-134269.

Park, S., & Lee, Y. (2021). Verification of Effectiveness of

Wearing Compression Pants in Wearable Robot Based

on Bio-signals. Journal of the Korean Society of

Clothing and Textiles, 45(2), 305-316.

Pinto-Fernandez, D., Torricelli, D., Sanchez-Villamanan,

M. D. C., Aller, F., Mombaur, K., Conti, R.,... Pons, J.

L. (2020). Performance Evaluation of Lower Limb

Exoskeletons: A Systematic Review. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, 28(7), 1573-1583.

Rizal, A., Hadiyoso, S., & Ramdani, A. Z. (2022). FPGA-

Based Implementation for Real-Time Epileptic EEG

Classification Using Hjorth Descriptor and KNN.

Electronics, 11(19), 3026

Song, S., & Collins, S. H. (2021). Optimizing Exoskeleton

Assistance for Faster Self-Selected Walking. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, 29, 786-795.

Ubeyli, E. D. (2008). Analysis of EEG signals by

combining eigenvector methods and multiclass support

vector machines. Computers in Biology and Medicine,

38, 14-22

Wang, W., Chen, J., Ding, J., Zhang, J., & Liu, J. (2022).

Improving Walking Economy With an Ankle

Exoskeleton Prior to Human-in-the-Loop

Optimization. Frontiers in Neurorobotics

, 15, 797147

Wickramasinghe, I., & Kalutarage, H. (2021). Naive Bayes:

applications, variations and vulnerabilities: a review of

literature with code snippets for implementation. Soft

Computing, 25, 2277-2293

Willett, F. R., Avansino, D. T., Hochberg, L. R.,

Henderson, J. M., & Shenoy, K. V. (2021). High-

performance brain-to-text communication via hand-

writing. Nature, 593, 249-254

Zhang, J., Fiers, P., Witte, K. A., Jackson, R. W.,

Poggensee, K. L., Atkeson, C. G.,... Collins, S. H.

(2017a). Human-in-the-loop optimization of

exoskeleton assistance during walking. Science,

356(6344), 1280-1284.

Zhang, J., Fiers, P., Witte, K. A., Jackson, R. W.,

Poggensee, K. L., Atkeson, C. G.,... Collins, S. H.

(2017b). Human-in-the-loop optimization of

exoskeleton assistance during walking. SCIENCE,

356(6344), 1280-1284.

Zhou, C., Yang, L., Liao, H., Liang, B., & Ye, X. (2021).

Ankle foot motion recognition based on wireless

wearable sEMG and acceleration sensors for smart

AFO. Sensors and Actuators A: Physical, 331.

Zhou, C., Yang, Z., Li, K., & Ye, X. (2022). Research and

Development of Ankle– Foot Orthoses: A Review.

sensors, 22(17), 6596

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

682