Design of BCI-Based Exoskeleton System for Knee Rehabilitation

Maryam Khoshkhooy Titkanlou

, Duc Thien Pham

and Roman Mouček

Department of Computer Science and Engineering, University of West Bohemia, 301 00 Plzen, Czech Republic

Keywords: Brain-Computer Interface (BCI), Electroencephalography (EEG), Exoskeleton, Knee, Lower Limb, Motor

Imagery (MI), Transfer Learning.

Abstract: Injuries of the lower limb, particularly the knee, usually require several months of rehabilitation. Exoskeletons

are great tools supporting the rehabilitation process; their research and suitable practical use are at the center

of interest of researchers and physiotherapists. This paper focuses on designing a brain-computer-interface

(BCI)-controlled exoskeleton for knee rehabilitation. It includes reviewing and selecting

electroencephalography (EEG) acquisition methods, BCI paradigms, current acquisition devices, signal

classification methods and techniques, and the target group of people for whom the exoskeleton will be

suitable. Finally, the preliminary proposal of the exoskeleton is provided.

1 INTRODUCTION

The number of people with lower limb movement

disorders due to aging and paralysis is increasing.

Exoskeletons can be a promising solution and a

useful, practical medical device in these cases;

Recently, exoskeletons have become a powerful tool

for the clinical rehabilitation of people with impaired

lower-limb function.

However, the proper design of exoskeletons is not

easy; exoskeletons should be lightweight, enabling

movements during rehabilitation on the one hand and

preventing health-hazardous movements on the other.

Another step in their improvements is introducing

active exoskeletons, i.e., exoskeletons that can be

controlled directly by impaired people and supported

with pneumatic control. This direct control should be

carried out remotely by using, for example, speech

commands or the human brain itself.

To design and implement such controllers,

researchers have recently used various biological

signals to control exoskeletons and other

neuroprosthetic devices. As one of the results, Brain-

Computer Interface (BCI) controllers based on

electroencephalographic (EEG) signals can

potentially (among others) bridge users’ need for

control and related rehabilitation devices

https://orcid.org/0000-0002-4139-6836

https://orcid.org/0000-0002-3037-5298

https://orcid.org/0000-0002-4665-8946

(exoskeletons), especially when the user needs to

rehabilitate the motor functions and the brain parts

responsible for movements in parallel.

BCIs are designed to decode intent by extracting it

from the human brain and its neural activity (Lin &

Lin, 2023). The main applications of BCIs have been

in communication with people in locked-in states and

just in rehabilitation, control of prosthetics, and

neurofeedback.

Specific protocols and paradigms need to be

chosen to implement an EEG-based BCI system for a

particular application. First, the user performs a

particular task (e.g., movement imagery or visual

task) (to learn) to modulate their brain activity while

EEG signals are recorded from the scalp. A neural

decoder for the paradigm is designed using the

recorded EEG as underlying (training) data.

Afterward, the user performs the task again, and the

neural decoder is used for BCI control (Orban et al.,

2022).

There are various experimental methods,

paradigms, and protocols for EEG data acquisition,

such as motor imagery (MI), active movement,

movement intention-active movements, assisted

movements, and electrical lower limb stimulation

(others are described later) to get suitable EEG data

for the control of exoskeleton. For example, MI

212

Khoshkhooy Titkanlou, M., Pham, D. and Mou

cek, R.

Design of BCI-Based Exoskeleton System for Knee Rehabilitation.

DOI: 10.5220/0012688800003699

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

In Proceedings of the 10th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2024), pages 212-220

ISBN: 978-989-758-700-9; ISSN: 2184-4984

allows users to control systems by imagining the

movements of their limbs, and the related EEG signal

is collected when the person imagines the movement.

The recorded EEG signal needs to be processed and

classified before it is used to actuate the exoskeleton,

i.e., the methods for lower limb movement detection

and classification need to be proposed, applied, and

validated.

This study investigates the experience and best

practices used to design and operate a successful

exoskeleton controlled by the human brain and to

look for BCI-based exoskeleton systems for lower-

limb and knee rehabilitation. Finally, the custom

proposal for such a system is shortly presented.

The paper is organized as follows. Section 2

reviews scientific papers on EEG-based control to

detect movement intention using various

experimental approaches, BCI paradigms, and EEG

signal classification methods. Also, current

companies' products are shortly presented. Section 3

presents our design proposal for this purpose. The

final section concludes the findings.

2 STATE-OF-THE-ART

We perform a search of articles in the field of BCI-

based exoskeleton system for knee rehabilitation. In

our search we used the general keywords such as

“knee”, “lower limb”, “EEG”, “exoskeleton”, “motor

imagery”, “transfer learning”, “deep learning”. Brain-

computer interface (BCI) is an emerging research

field that creates a real-time bidirectional connection

between the human brain and a computer/output

device. It is a communication tool for patients with

neuromotor disorders, spinal cord injuries, or

amputations (Tariq et al., 2018) (Lebedev &

Nicolelis, 2017). Among its various applications, the

most popular one is neurorehabilitation, which

involves sensory feedback and the use of brain-

controlled biomedical devices, e.g. exoskeletons.

Neurorehabilitation, a critical component of the

recovery process for those with neurological

impairments, is being revolutionized by integrating

exoskeletons. These wearable robotic devices have

shown significant potential in enhancing mobility,

functional independence, and overall quality of life

for people suffering, e.g., from spinal cord injury,

stroke, and multiple sclerosis. Moreover, the lower

limb exoskeletons are also effective tools for

clinically rehabilitating people with impaired lower

limb function due to injury.

In the BCI community, many BCI systems have

utilized the classification of imaginary upper limb

movements, e.g. (Paredes-Acuna et al., 2024) (Liao et

al., 2014) to generate different commands for

controlling devices, including robots (Jeon et al.,

2024). Only a few studies addressed the MI problem

of the lower limb, and these studies were all focused

on the imagination of brisk foot movement (ankle

dorsiflexion) (Xu et al., 2014). The main reason is

that the left and right foot representation areas in the

sensorimotor cortex are very close to each other and

located deeply within the interhemispheric fissure.

The following parts review papers based on

various experimental methodologies for data

acquisition, BCI paradigms, products of BCI

companies, and classification methods. Their

interesting characteristics and results (such as used

protocols, tasks, EEG channels, preprocessing and

feature extraction methods, and accuracies) are

summarized in Table 1 (EEG-based control for lower

limb movements). The summary of investigations

utilizing transfer learning is given in Table 2

(Summary of transfer learning for MI Classification

using EEG signal).

2.1 Experimental Methodologies

We can classify all experimental methodologies used

to record EEG signals for lower limbs into the

following types:

In most cases, alpha power (8-13 Hz) is

suppressed, while beta power (13-30 Hz) increases,

when an individual is executing tasks that require

concentration which is highly related to motor

imagery.

Active movement-based (AcM) BCIs can work

well for individuals with sufficient residual control

over their knee joints. By using brain signals related

to particular movements, AcM BCI allows users to

manipulate external devices in real time. Tortora et al.

(Tortora et al., 2023) recorded EEG and

electromyographic (EMG) activity from ten healthy

volunteers walking with an exoskeleton. Choi et al.

(Choi et al., 2020) recorded EEG signals from 10

healthy volunteers. All volunteers were right-handed

males with no history of neurological disorders. The

volunteers had to sit and walk while making energetic

movements. The patients were given visual cues

when it was time to do the movement. Ten healthy

subjects participated in the offline and online

sessions, and the average classification accuracy was

more than 80% for both sessions.

Motor imagery (MI) is viable if the user can

imagine movements. It's a non-invasive approach that

doesn't require physical movement, making it suitable

for users with various mobility levels. Hsu et al. (Hsu

Design of BCI-Based Exoskeleton System for Knee Rehabilitation

213

et al., 2017) recorded EEG signals from eight healthy

volunteers. The volunteers' tasks included both left

and right stepping. Because a screen was used for

visual stimulation, electrooculography (EOG) was

employed as an extra sensor.

Event-related desynchronization (ERD) reflects a

decrease in oscillatory activity related to internally or

externally paced events. The increase in rhythmic

activity is called event-related synchronization

(ERS). Event-Related Desynchronization and Event-

Related Synchronization (ERD/ERS) EEG signals

were recorded from 14 healthy participants by Tariq

et al. (Tariq et al., 2019). During the experiment,

participants completed MI tasks while seated.

Jeong et al. (Jeong et al., 2022) recorded two

lower-limb MIs (gait and sit-down) and resting EEG

data from five healthy subjects. The subjects were

asked to stand comfortably in front of the monitor and

start the MI task when ready through a mouse click.

Then, the subjects performed two MI tasks related to

the lower limb and rested for five seconds according

to the monitor's visual cues. Roy and Bhaumik (Roy

& Bhaumik, 2022) recorded EEG signals from three

participants. The protocol consisted of four MI-

related tasks: the imagination of left hand (L), right

hand (R), foot (F), and tongue (T) movement.

Combining both MI and AcMs can provide more

flexibility. Users can imagine knee movements when

their physical capabilities are limited or actively

move when they can. Lins (Lin & Lin, 2023) recorded

EEG signals from eight healthy subjects for MI tasks

at rest and during walking. Gordleeva et al.

(Gordleeva et al., 2020) recorded EEG signals from

eight healthy volunteers. EEG and EMG signals for a

leg lift movement were acquired. AcMs and MI were

the tasks completed. EMG sensors were also

employed to provide feedback to the exoskeleton

control system for the lower limb. Li et al. (Li et al.,

2022) recorded EEG signals and sEMG signals

controlled by the participants’ brains on the arms of

two healthy subjects. The task performed was based

on MI and AcM.

Another methodology is based on Motor imagery,

Active movements, and Attempted movements

(AtMs): Unlike active movement, which depends on

utilizing brain signals connected to actual physical

actions for real-time interaction, attempted movement

focuses on interpreting neural signals related to

individuals' intentions to move, allowing control of

external devices without physical execution.

Jochumsen et al. (Jochumsen et al., 2015) recorded

EEG signals from twelve healthy subjects and six

stroke patients with lower limb paresis. The subject

was seated in a comfortable chair with the right foot

(or the affected foot) attached to a foot pedal where a

force transducer was set up. The healthy subjects

performed the two tasks with Motor Execution (ME)

and Motor Imagery (MI), while the stroke patients

were asked to attempt the movements.

Movement intention - Active movements:

Movement intention (like attempted movement)

refers to the mental state in which an individual

intends to carry out a specific action, even before the

actual execution. Movement intention-based BCIs

benefit users who want the exoskeleton to respond to

their intentions even before visible movements occur.

Rea et al. (Rea et al., 2014) recorded EEG signals

from seven right-handed patients with chronic stroke.

The subjects were seated during the experiment and

performed movements with a foot pedal. The authors

employed additional EMG sensors during the tasks.

Assisted movement benefits users with severe

mobility impairments; an exoskeleton with BCI-

controlled assisted movements can be the best option.

This method involves integrating BCI technology to

improve physical movements, providing people with

assistance or control over external devices to augment

their motor functions. Qiu et al. (Qiu et al., 2015)

recorded Event-Related Desynchronization (ERD)

EEG signals from 12 healthy volunteers and a stroke

patient with hemiplegia. The tasks performed were

right-leg lifts.

Electrical lower-limb stimulation is suitable if the

user has complete paralysis but still wants to engage

in knee movements; electrical lower-limb stimulation

controlled by a BCI may be the most suitable option.

Hauck et al. (Hauck et al., 2006) recorded EEG

signals from six healthy right-handed volunteers.

Furthermore, Magnetic Resonance Imaging (MRI)

was obtained from five volunteers for data recording.

Subjects were lying down, and low-amperage

electrical stimulation was applied to the peroneal,

proximal, and distal tibial nerves. Sensors for

electrooculography (EOG) were also employed.

2.2 BCI Paradigms

BCIs can be divided into two main categories:

invasive and non-invasive (Sitaram et al., 2007).

Most of the EEG-based BCI systems rely on the

following paradigms: ERD associated with motor

imagery (MI), event-related potentials (ERPs) based

on the P300 or other event-related components,

steady-state visual evoked potentials (SSVEPs),

auditory steady-state responses (ASSRs), slow

cortical potentials (SCPs), sensorimotor rhythm

(SMR), and various hybrid systems based on more

than one input signal (Orban et al., 2022).

ICT4AWE 2024 - 10th International Conference on Information and Communication Technologies for Ageing Well and e-Health

214

ERD is widely used in MI tasks for motor

rehabilitation and control of prosthetic limbs,

allowing users to control devices or perform actions

by imagining specific movements. ERP BCIs are

used for communication, cognitive, and clinical

applications. SSVEP BCIs are often used in

applications requiring high information transfer rates,

such as gaming and spelling systems.

ASSRs are less common than visual-based BCIs

but can be used for auditory communication and

spatial tasks. SCPs are primarily used for

neurofeedback and cognitive regulation, such as

improving attention and relaxation.

SMR requires people to use mental strategies or

MI to enable motor execution (ME). For subjects with

motor disabilities, the thought of movement can

suppress EEG rhythm, leading to desynchronization,

resulting in movement initiation. By harnessing

neuroplasticity, MI can enhance motor learning

(Müller-Putz et al., 2005). With both MI and ME

derived from sensorimotor areas such as the primary

motor area, supplementary motor area, and premotor

cortex, SMR can be manipulated to help disabled

people towards rehabilitation. The SMR paradigm

has been one of the most promising paradigms used

by people with tetraplegia, spinal cord injury, and

amyotrophic lateral sclerosis (ALS) (Kawala-

Sterniuk et al., 2021).

Hybrid BCIs combine multiple input signals, such

as EEG, ECoG, EMG, other physiological measure-

ments, and various paradigms to enhance overall BCI

performance and functionality. They are used in

applications where high accuracy, versatility, and

adaptability are needed, such as advanced prosthetic

control and complex communication systems.

The primary EEG-based BCI paradigms for lower

limb rehabilitation are ERD associated with MI,

SMR, and Hybrid BCIs (combining EEG with other

modalities).

2.3 BCI Products

EEG acquisition systems and related BCI systems

have also become popular during the last few years;

many companies were founded to produce simpler

and cheaper BCI systems for ordinary users, but

qualitatively compared to those intended for

fundamental research on the brain's functioning.

The most frequently used EEG (BCI) headsets are

delivered by the following companies: Emotiv Inc.

(San Francisco, CA, USA), Ant Neuro (Hengelo,

Netherlands), Cognionics (San Diego, CA, USA),

Neurosky Inc. (San Jose, CA, USA), OpenBCI

(Brooklyn, NY, USA), InteraXon (Toronto, Canada),

g.tec (Schiedlberg, Austria), and CREmedical

(Kingston, RI, USA). The products continuously

improve signal acquisition quality, wearing comfort,

raw EEG signal preprocessing, or accompanying

software tools. The critical feature for further EEG

signal processing is the accessibility of raw EEG data.

In comparison, g.tec provided a stronger and cleaner

signal than Emotiv Inc. The biggest advantage of

Neurosky Inc. products is a low, competitive price

and ease of use. The OpenBCI provided extremely

similar EEG results to those obtained with the g.tec

device, and the medical-grade equipment performed

marginally better than the consumer-grade one and

OpenBCI gave very close EEG readings to those

obtained with the g.tec device (Kawala-Sterniuk et

al., 2021). Most of the clinical-quality EEG data for

the BCI applications are gathered using the following

clinical-grade amplifiers which are popular mostly

due to their price, availability, and the high quality-

signals they provide: g.tec amplifiers, Porti7 (TMSI),

Nuamp amplifier, BrainAmp128DC, and

BioNomadix amplifier (Biopac).

Clinical-level (medical devices) EEG equipment

is also popular in numerous BCI-related applications.

In many cases, the g.tec (Kuś et al., 2013) amplifiers

are used, e.g., BCI systems dedicated to controlling a

neuroprosthesis (Tung et al., 2013). Another popular

clinical-level device is Porti7 from the TMSI

company, which was applied for an SSVEP-BCI

system, where the authors tried to find the most

appropriate SSVEP frequencies (Onose et al., 2012).

The neuroscan device Nuamp was applied for BCI-

based post-stroke patients’ rehabilitation (Fazli et al.,

2009). BrainAmp128DC was used in studies (Katona

& Kovari, 2018) (Fazli et al., 2009) to gather EEG-

based robotic arm control data. In (Katona & Kovari,

2018), the authors compared the inexpensive

Neurosky’s Mindwave device with Biopac’s

BioNomadix amplifier, and the obtained results

proved the similar quality of the recorded data. Based

on a new research, market leaders for medical

exoskeletons included Ekso Bionics Holdings,

Rewalk Robotics, Cyberdyne, Bionik Laboratories,

Bioness Inc. (Exoskeleton Market - Size, Growth &

Trends, n.d.).

In the next section devices related to the BCI

companies and parameters for choosing suitable

devices are discussed.

2.4 EEG-Related Devices

Several important parameters should be considered

when choosing EEG devices for MI in rehabilitation

to ensure that the chosen device fits the unique

Design of BCI-Based Exoskeleton System for Knee Rehabilitation

215

requirements and objectives of the rehabilitation

program. Some crucial selection standards involve

wireless connectivity; devices with wireless

capabilities increase the mobility and flexibility of

users during rehabilitation exercises. The placement

and number of electrodes ensure the device captures

relevant brain activity associated with MI and

accurate monitoring. Researchers frequently place

electrodes over the sensorimotor cortex (C3 and C4)

in the context of motor imagery BCIs because these

regions are directly involved in the mental simulation

of movement. There are equivalent changes in brain

activity in the sensorimotor cortex when an individual

performs motor imagery, such as imagining moving

their left hand. These electrical potential changes

corresponding to motor imagery can be recorded by

EEG electrodes located at C3, C4 and Cz.

Higher sampling rate and data resolution help to

provide more precise and thorough brain activity

tracking if necessary. Lightweight and portable

devices are ideal for rehabilitation environments

where people must move freely. Long battery life is

essential for continuous monitoring sessions without

frequent interruptions for recharging. The overall cost

of the EEG device, including any additional

accessories, software licenses, or maintenance fees,

should be balanced with the available budget for the

rehabilitation program.

Emotiv company offers solutions for BCI

applications, including MI tasks. It is well-known for

its Emotiv EPOC+ and Emotiv Insight EEG headsets.

The Emotiv EPOC headset is straightforward to use

and does not require any particular scalp preparation.

NeuroSky provides a small wireless MindWave

Mobile EEG headset, which can be used for BCI and

neurofeedback applications. The biggest advantage of

NeuroSky products is a low, competitive price. g.tec

is known for its excellent data resolution and

adaptable features. One of the most popular

clinically-approved professional EEG systems is

g.USBAMP from the g.tec company. It is a cheap

device (ca. 25 USD), providing excellent data quality.

OpenBCI is renowned for offering configurable and

open-source EEG platforms. OpenBCI Cyton and

Ganglion are two of their popular offerings among

developers and researchers.

2.5 Deep Learning-Based Approaches

Applied in MI Classification

Several lower limb movement classification models

have recently emerged, utilizing machine learning

(ML) and deep learning (DL) techniques for EEG

data processing. Hsu et al. (Hsu et al., 2017)

employed a Fuzzy SVM (FSVM) approach to classify

imagined lower-limb stepping movements and create

a resilient MI classifier. They utilized data from nine

EEG channels and electrooculography (EOG)

signals. The highest performance, with a high average

classification accuracy across eight subjects (86.25%

in single-trial analysis), was attained using a filter

bank common spatial pattern (FBCSP) and FSVMb

combination.

Gordleeva et al. (Gordleeva et al., 2020) proposed

a multimodal human-machine interface (mHMI) that

integrates EEG and EMG modalities for real-time

control of a lower-limb exoskeleton. The

classification and control system based on linear

discriminant analysis (LDA) achieved successful

movement prediction and differentiation (81.5% ±

14.9%) using the combined EEG and EMG signals.

In another study by Roy et al. (Roy et al., 2022),

LDA was utilized to classify the walking MI task,

achieving an accuracy of 98.67%. This investigation

involved data from a dataset comprising 32 EEG

channels collected from five healthy subjects.

Similarly, Roy & Bhaumik (Roy & Bhaumik,

2022) employed LDA to classify four MI tasks,

including left hand (L), right hand (R), foot (F), and

tongue (T) movements, using data from 3 EEG

channels (C3, Cz, C4). Their study demonstrated a

classification accuracy of 88.89%.

Tortora et al. (Tortora et al., 2023) employed a

Convolutional Neural Network-Long Short-Term

Memory (CNN-LSTM) hybrid model to classify the

walking MI task. They utilized data from 38 EEG

channels in conjunction with EMG and inertial

measurement unit (IMU) information gathered from

10 healthy subjects. Their approach achieved a

classification accuracy of 89.32% ± 4.65%. Table 1

shows EEG-based control for lower limb movements

with MI and Active Movement using ML and DL for

classification.

Deep Neural Networks (DNNs) have emerged as

game changers in ML and DL, capable of tackling

complex tasks with remarkable accuracy. However,

training DNNs from scratch can be computationally

intensive and data-hungry, limiting their practical

utility. This is where transfer learning (TL), especially

using pre-trained networks, comes into play.

TL is a technique that leverages the knowledge

gained from one task and applies it to a different,

often related task. TL is a promising approach to

address these challenges by transferring knowledge

from related tasks to improve learning ability. TL can

help improve the performance of decoding models

across subjects/sessions and reduce the calibration

time of BCI systems.

ICT4AWE 2024 - 10th International Conference on Information and Communication Technologies for Ageing Well and e-Health

216

Table 1: EEG-based control for lower limb movements.

Study Subjects Protocol Task EEG channel

Pre

Processing

Feature

extraction

Method

Accuracy

(%)

(Hsu et al.,

2017)

8 healthy

subjects

MI - Active

Movement

Left/Right

stepping

9 channels (FC3,

FC4, FCz, C3,

C4, Cz, CP3,

CP4, and CPz)

Filter-bank common spatial

Pattern (FB-CSP)

Fuzzy SVM

(FSVM)

86.25

(Tariq et al.,

2019)

5 healthy

subjects

Kinaesthetic Motor

Imagery (KMI)

left/right knee

extension

19 channels FB-CSP

Logistic Regression

(Logreg)

70.0 ± 2.85

(Gordleeva et

al., 2020)

8 healthy

subjects

MI - Active

Movement

Leg lift

7 channels (C5,

C3, C1, Cz, C2,

C4, C6)

Bandpass filters CSP LDA 81.5 ± 14.9

(Roy et al.,

2022)

5 healthy

subjects

MI Walk

channels

Band pass filter

Cross-

correlation and

Spectral

entropy

LDA 98.67

(Jeong et al.,

2022)

5 healthy

subjects

gait and

sit-down

channels

Band pass filter

Dual-domain

CNN

Dual-domain CNN

based subject-

transfer approach

66.57 ± 7.33

(Roy &

Bhaumik,

2022)

3 subjects MI

left hand (L),

right

hand (R), foot

(F) and tongue

(T) movement

3 channels (C3,

Cz, C4)

Band pass filter

Cross-

correlation and

Wavelet

Energy

LDA 88.89

(Tortora et

al., 2023)

10 healthy

subjects

Active Movement Walk

channels

High-pass,

notch and low-

ass filtere

PSD CNN-LSTM 89.32 ± 4.65

(Lin & Lin,

2023)

8 healthy

subjects

MI-ME (motor

execution)

Walk & Stand

Channels (FP1,

FP2, C3, Cz, C4,

CP3, CPz, CP4)

Band pass filter

CSP, PSD,

DWT+AR

SVM 83.09

(Li et al.,

2022)

2 healthy

subjects

Active Movement Walk & Stand

32 Channels of

EEG

Fifth-order

Butterworth

filter and

Notch filte

CSP ICA 99.0

Table 2: Summary of transfer learning for MI Classification using EEG signal.

Study Subjects Protocol Task EEG channel

Pre

Processing

Feature

extraction

Method

Accuracy

(%)

(Zheng et al.

2020)

10 healthy

subjects

MI Left/right hand

8 channels (Cz,

C3, C4, CP1, CP2,

Pz, P3, and P4)

CSP,

PSD

classical transfer

learning algorithm

91.6 ± 2.8

(Liang., &

Ma, Y.

(2020).

12 subjects MI

right

hand and both

feet

13 channels band-pass filtered

balanced

distribution

adaptation (BDA)

multi-source

fusion transfer

learning (MFTL)

71.89

(Kant et al.,

2020)

a healthy

female

left/right hand

movement

3 channels (C3, Cz

and C4)

band-pass filtered

Continuous

Wavelet

Transform (CWT)

VGG19 95.71

(Zhang et al.,

2021)

9 healthy

subjects

left hand, right

hand, feet, and

tongue

22 channels OVR-FBCSP HDNN-TL 81

(Zhang et al.,

2021)

54 healthy

subjects

grasping with

the hand

62 channels

Chebyshev type-I

filter

- Deep CNN 84.19 ± 9.98

(Mattioli et

al., 2021)

109

participants

4 tasks and 14

experimental

runs

64 channels - - 1D CNN 99.46

(Cai et al.,

2022)

- Dataset 1

- Dataset 2a

left hand, right

hand, foot

- 59 channels

- 22 channels

band-pass filter

symmetric

positive definite

(SPD) and

Grassmann

manifold

embedded transfer

learning (METL)

- 83.14

- 76.00

(Khademi et

al., 2022)

9 healthy

subjects

left hand, right

hand, feet, and

tongue

22 channels

spatial and

frequency domains

CWT

Inception-v3 and

LSTM

Design of BCI-Based Exoskeleton System for Knee Rehabilitation

217

Pre-trained CNN networks, such as VGGNet,

AlexNet, ResNet, Inception, and GoogleNet, among

others, represent the cornerstone of transfer learning.

These networks are initially trained on extensive

datasets for general tasks and serve as the foundation

for our research, which focuses on classifying lower-

limb MI using EEG signals. Utilizing pre-trained

CNN networks, a form of TL based on model

parameters, for lower limb movement classification

using EEG signal offers several significant

advantages and compelling reasons, such as

complexity of EEG data, data efficiency,

generalization, model performance, reduction in

overfitting, knowledge transfer, reduced

computational resource, and robustness to noise.

Table 2 provides a comprehensive summary of recent

research that has successfully harnessed the power of

TL to achieve high-performance MI classification.

Kant et al. (Kant et al., 2020) proposed a

combination of Continuous Wavelet Transform

(CWT) along with deep learning-based transfer

learning (pre-trained CNN like VGG19) using three

EEG channels (C3, Cz, C4) for MI Classification for

BCI. The results of the method have been compared

to earlier works on the same dataset, and a promising

validation accuracy of 95.71% is achieved in their

investigation.

Khademi et al. (Khademi et al., 2022) employed a

transfer learning-based CNN (ResNet-50 and

Inception-v3) and LSTM hybrid deep learning model

to classify MI EEG signals. Their model produced

impressive results, achieving the highest accuracy of

92% and a Kappa value of 88% for the hybrid neural

network featuring Inception-v3.

3 BCI-CONTROLLED

EXOSKELETON FOR KNEE

REHABILITATION

Summarizing the literature, BCI paradigms, EEG

acquisition methods, current BCI products,

classification and transfer learning methods, and

approaches, the general usefulness of the BCI

exoskeleton and possible target groups, we have

decided to propose the BCI-controlled exoskeleton

for knee rehabilitation.

This exoskeleton emerges as a promising solution

for post-knee injury rehabilitation, particularly in

cases without neurological diseases. Its lightweight

design facilitates permissible movements during

rehabilitation, and its active functionality, driven by

both brain-controlled and pneumatic systems,

positions it as an effective tool, particularly for bed-

based rehabilitation scenarios. The BCI system will

be based on two paradigms: processing and

evaluating basic brain frequencies (distinguishing

between attention and relaxation states) and motor

imagery. As we move forward, it is crucial for

individuals to recognize the significance of

rehabilitation, emphasizing the role of attention and

motor imagery in optimizing the efficacy of the

exoskeleton and the overall recovery process.

The basic EEG acquisition device used will be

based on Neurosky technology. We also aim to define

MI tasks with the SMR paradigm for implementing

experiments on healthy subjects using the OpenBCI

Cyton device available in our laboratory. Regarding

the classification method, we intend to use pre-trained

CNN networks (such as VGG19) to classify lower

limb movements. This interface leverages pre-trained

CNN networks to extract shared features from EEG

signals, thus enhancing the performance of lower

limb movement classification in the human-

exoskeleton interface. The benefits of transfer

learning will be also utilized. Currently, the

microcontroller for running the BCI part (i.e.

collecting the EEG signal from the EEG acquisition

device and running the classification methods) is

designed, created, and tested in the laboratory.

4 CONCLUSION

This paper reviewed EEG current literature,

acquisition methods, BCI paradigms, current EEG

acquisition devices, and EEG signal classification

methods and techniques to design and implement a

BCI-controlled exoskeleton for knee rehabilitation.

This preliminary design of such a system was

presented. When the BCI part is ready for lab testing,

the future work involves mainly building the interface

for the exoskeleton and performing experiments

when the BCI system and exoskeleton are integrated.

ACKNOWLEDGEMENTS

This work was supported by the University specific

research project SGS-2022-016 Advanced Methods

of Data Processing and Analysis (project SGS-2022-

016) and SGS-2022-015 New Methods for Medical,

Spatial and Communication Data (project SGS-2022-

015).

ICT4AWE 2024 - 10th International Conference on Information and Communication Technologies for Ageing Well and e-Health

218

REFERENCES

Lin, C.-J., & Lin, C.-H. (2023). Classification of EEG Signals

Using a Common Spatial Pattern Based Motor-Imagery

for a Lower-limb Rehabilitation Exoskeleton. IEEE

EUROCON 2023-20th International Conference on

Smart Technologies, 764–769. https://ieeexplore.

ieee.org/abstract/document/10198960/.

Orban, M., Elsamanty, M., Guo, K., Zhang, S., & Yang, H.

(2022). A Review of Brain Activity and EEG-Based

Brain–Computer Interfaces for Rehabilitation

Application. Bioengineering, 9(12), 768. https://doi.org/

10.3390/bioengineering9120768.

Tariq, M., Trivailo, P. M., & Simic, M. (2018). EEG-based

BCI control schemes for lower-limb assistive-robots.

Frontiers in Human Neuroscience, 12, 312.

Lebedev, M. A., & Nicolelis, M. A. L. (2017). Brain-

Machine Interfaces: From Basic Science to

Neuroprostheses and Neurorehabilitation. Physiological

Reviews, 97(2), 767–837. https://doi.org/10.1152/

physrev.00027.2016.

Paredes-Acuna, N., Utpadel-Fischler, D., Ding, K., Thakor,

N. V., & Cheng, G. (2024). Upper limb intention tremor

assessment: Opportunities and challenges in wearable

technology. Journal of NeuroEngineering and

Rehabilitation, 21(1), 8. https://doi.org/10.1186/s129 84-

023-01302-9.

Liao, K., Xiao, R., Gonzalez, J., & Ding, L. (2014). Decoding

individual finger movements from one hand using human

EEG signals. PloS One, 9(1), e85192.

Jeon, S. Y., Ki, M., & Shin, J.-H. (2024). Resistive versus

active assisted robotic training for the upper limb after a

stroke: A randomized controlled study. Annals of

Physical and Rehabilitation Medicine, 67(1), 101789.

https://doi.org/10.1016/j.rehab.2023.101789.

Choi, J., Kim, K. T., Jeong, J. H., Kim, L., Lee, S. J., & Kim,

H. (2020). Developing a motor imagery-based real-time

asynchronous hybrid BCI controller for a lower-limb

exoskeleton. Sensors, 20(24), 7309.

Fatourechi, M., Ward, R. K., & Birch, G. E. (2007). A self-

paced brain–computer interface system with a low false

positive rate. Journal of Neural Engineering, 5(1), 9.

Fazli, S., Danóczy, M., Popescu, F., Blankertz, B., & Müller,

K.-R. (2009). Using Rest Class and Control Paradigms

for Brain Computer Interfacing. In J. Cabestany, F.

Sandoval, A. Prieto, & J. M. Corchado (Eds.), Bio-

Inspired Systems: Computational and Ambient

Intelligence (Vol. 5517, pp. 651–665). Springer Berlin

Heidelberg. https://doi.org/10.1007/ 978-3-642-02478-

8_82.

Gardner, A. D., Potgieter, J., & Noble, F. K. (2017). A review

of commercially available exoskeletons’ capabilities.

2017 24th International Conference on Mechatronics

and Machine Vision in Practice (M2VIP), 1–5.

https://ieeexplore.ieee.org/abstract/ document/8211470/.

Gordleeva, S. Y., Lobov, S. A., Grigorev, N. A., Savosenkov,

A. O., Shamshin, M. O., Lukoyanov, M. V., Khoruzhko,

M. A., & Kazantsev, V. B. (2020). Real-time EEG–EMG

human–machine interface-based control system for a

lower-limb exoskeleton. IEEE Access, 8, 84070–84081.

G.tec medical engineering GmbH | Brain-Computer

Interfaces & Neurotechnology

. (n.d.). Retrieved January

20, 2024, from https://www.gtec.at/.

Hauck, M., Baumgärtner, U., Hille, E., Hille, S., Lorenz, J.,

& Quante, M. (2006). Evidence for early activation of

primary motor cortex and SMA after electrical lower

limb stimulation using EEG source reconstruction. Brain

Research, 1125(1), 17–25.

Hsu, W.-C., Lin, L.-F., Chou, C.-W., Hsiao, Y.-T., & Liu, Y.-

H. (2017). EEG Classification of Imaginary Lower Limb

Stepping Movements Based on Fuzzy Support Vector

Machine with Kernel-Induced Membership Function.

International Journal of Fuzzy Systems, 19(2), 566–579.

https://doi.org/10.1007/s40815-016-0259-9.

Jeong, J.-H., Kim, K.-T., Lee, S. J., Kim, D.-J., & Kim, H.

(2022). CNN-based Subject-Transfer Approach for

Training Minimized Lower-Limb MI-BCIs. 2022 10th

International Winter Conference on Brain-Computer

Interface (BCI), 1–4. https://ieeexplore.ieee.org/

abstract/document/9734910/.

Jochumsen, M., Khan Niazi, I., Samran Navid, M., Nabeel

Anwar, M., Farina, D., & Dremstrup, K. (2015). Online

multi-class brain-computer interface for detection and

classification of lower limb movement intentions and

kinetics for stroke rehabilitation. Brain-Computer

Interfaces, 2(4), 202–210. https://doi.org/10.1080/

2326263X.2015.1114978

Exoskeleton Market—Size, Growth & Trends. (n.d.).

Retrieved January 25, 2024, from https://www.

mordorintelligence.com/industry-reports/exoskeleton-

market.

Katona, J., & Kovari, A. (2018). The evaluation of bci and

pebl-based attention tests. Acta Polytechnica Hungarica,

15(3), 225–249.

Kawala-Sterniuk, A., Browarska, N., Al-Bakri, A., Pelc, M.,

Zygarlicki, J., Sidikova, M., Martinek, R., &

Gorzelanczyk, E. J. (2021). Summary of over Fifty Years

with Brain-Computer Interfaces-A Review. Brain

sciences, 11(1), 43. https://doi.org/10.3390/brainsci1101

0043.

Kuś, R., Duszyk, A., Milanowski, P., \Labęcki, M.,

Bierzyńska, M., Radzikowska, Z., Michalska, M.,

Żygierewicz, J., Suffczyński, P., & Durka, P. J. (2013).

On the quantification of SSVEP frequency responses in

human EEG in realistic BCI conditions. PloS One, 8(10),

e77536.

Lebedev, M. A., & Nicolelis, M. A. L. (2017). Brain-

Machine Interfaces: From Basic Science to

Neuroprostheses and Neurorehabilitation. Physiological

Reviews, 97(2), 767–837. https://doi.org/10.1152/

physrev.00027.2016.

Li, W., Shao, K., Zhu, C., Ma, Y., Cao, W., Yin, M., Yang,

L., Luo, M., & Wu, X. (2022). Preliminary study of

online real-time control system for lower extremity

exoskeletons based on EEG and sEMG fusion. 2022

IEEE International Conference on Robotics and

Biomimetics (ROBIO), 1689–1694. https://ieeexplo

re.ieee.org/abstract/document/10011813/.

Liao, K., Xiao, R., Gonzalez, J., & Ding, L. (2014). Decoding

individual finger movements from one hand using human

Design of BCI-Based Exoskeleton System for Knee Rehabilitation

219

EEG signals. PloS One, 9(1), e85192.

Lin, C.-J., & Lin, C.-H. (2023). Classification of EEG Signals

Using a Common Spatial Pattern Based Motor-Imagery

for a Lower-limb Rehabilitation Exoskeleton. IEEE

EUROCON 2023-20th International Conference on

Smart Technologies, 764–769. https://ieeexplo

re.ieee.org/abstract/document/10198960/.

Müller-Putz, G. R., Scherer, R., Pfurtscheller, G., & Rupp, R.

(2005). EEG-based neuroprosthesis control: A step

towards clinical practice. Neuroscience Letters, 382(1–

2), 169–174.

Onose, G., Grozea, C., Anghelescu, A., Daia, C., Sinescu, C.

J., Ciurea, A. V., Spircu, T., Mirea, A., Andone, I., &

Spânu, A. (2012). On the feasibility of using motor

imagery EEG-based brain–computer interface in chronic

tetraplegics for assistive robotic arm control: A clinical

test and long-term post-trial follow-up. Spinal Cord,

50(8), 599–608.

[PDF] Efficiency evaluation of external environments

control using bio-signals | Semantic Scholar. (n.d.).

Retrieved January 20, 2024, from https://www.seman

ticscholar.org/paper/Efficiency-evaluation-of-external-

environments-Kawala-

Janik/a213cb6e27f004e20406747730e1df6db6b24dee.

Qiu, S., Yi, W., Xu, J., Qi, H., Du, J., Wang, C., He, F., &

Ming, D. (2015). Event-related beta EEG changes during

active, passive movement and functional electrical

stimulation of the lower limb. IEEE Transactions on

Neural Systems and Rehabilitation Engineering, 24(2),

283–290.

Rea, M., Rana, M., Lugato, N., Terekhin, P., Gizzi, L., Brötz,

D., Fallgatter, A., Birbaumer, N., Sitaram, R., & Caria,

A. (2014). Lower Limb Movement Preparation in

Chronic Stroke: A Pilot Study Toward an fNIRS-BCI for

Gait Rehabilitation. Neurorehabilitation and Neural

Repair, 28(6), 564–575. https://doi.org/10.1177/15459

68313520410.

Roy, G., & Bhaumik, S. (2022). Classification of MI EEG

Signal Using Minimum Set of Channels to Control a

Lower Limb Assistive Device. Journal of The Institution

of Engineers (India): Series B. https://doi.org/10.1007/

s40031-022-00783-x.

Roy, G., Bhoi, A. K., Das, S., & Bhaumik, S. (2022). Cross-

correlated spectral entropy-based classification of EEG

motor imagery signal for triggering lower limb

exoskeleton. Signal, Image and Video Processing, 16(7),

1831–1839. https://doi.org/10.1007/s11760-022-02142-

Sitaram, R., Caria, A., Veit, R., Gaber, T., Rota, G., Kuebler,

A., & Birbaumer, N. (2007). FMRI brain-computer

interface: A tool for neuroscientific research and

treatment. Computational Intelligence and

Neuroscience, 2007. https://www.hindawi.com/journals/

cin/2007/025487/abs/.

Tariq, M., Trivailo, P. M., & Simic, M. (2018). EEG-based

BCI control schemes for lower-limb assistive-robots.

Frontiers in Human Neuroscience, 12, 312.

Tariq, M., Trivailo, P. M., & Simic, M. (2019). Classification

of left and right knee extension motor imagery using

common spatial pattern for BCI applications.

Procedia

Computer Science, 159, 2598–2606.

Tortora, S., Tonin, L., Sieghartsleitner, S., Ortner, R., Guger,

C., Lennon, O., Coyle, D., Menegatti, E., & Del Felice,

A. (2023). Effect of lower limb exoskeleton on the

modulation of neural activity and gait classification.

IEEE Transactions on Neural Systems and

Rehabilitation Engineering. https://ieeexplore.ieee.org/

abstract/document/10177994/.

Tung, S. W., Guan, C., Ang, K. K., Phua, K. S., Wang, C.,

Zhao, L., Teo, W. P., & Chew, E. (2013). Motor imagery

BCI for upper limb stroke rehabilitation: An evaluation

of the EEG recordings using coherence analysis. 2013

35th Annual International Conference of the IEEE

Engineering in Medicine and Biology Society (EMBC),

261–264. https://doi.org/10.1109/EMBC.2013.6609487.

Xu, R., Jiang, N., Mrachacz-Kersting, N., Lin, C., Prieto, G.

A., Moreno, J. C., Pons, J. L., Dremstrup, K., & Farina,

D. (2014). A closed-loop brain–computer interface

triggering an active ankle–foot orthosis for inducing

cortical neural plasticity. IEEE Transactions on

Biomedical Engineering, 61(7), 2092–2101.

Zheng, M., Yang, B., & Xie, Y. (2020). EEG classification

across sessions and across subjects through transfer

learning in motor imagery-based brain-machine interface

system. Medical & biological engineering & computing,

58, 1515-1528.

Liang, Y., & Ma, Y. (2020). Calibrating EEG features in

motor imagery classification tasks with a small amount

of current data using multisource fusion transfer learning.

Biomedical Signal Processing and Control, 62, 102101.

Kant, P., Laskar, S. H., Hazarika, J., & Mahamune, R. (2020).

CWT based transfer learning for motor imagery

classification for brain computer interfaces. Journal of

Neuroscience Methods, 345, 108886.

Zhang, R., Zong, Q., Dou, L., Zhao, X., Tang, Y., & Li, Z.

(2021). Hybrid deep neural network using transfer

learning for EEG motor imagery decoding. Biomedical

Signal Processing and Control, 63, 102144.

Zhang, K., Robinson, N., Lee, S. W., & Guan, C. (2021).

Adaptive transfer learning for EEG motor imagery

classification with deep convolutional neural network.

Neural Networks, 136, 1-10.

Mattioli, F., Porcaro, C., & Baldassarre, G. (2022). A 1D

CNN for high accuracy classification and transfer

learning in motor imagery EEG-based brain-computer

interface. Journal of Neural Engineering, 18(6), 066053.

Cai, Y., She, Q., Ji, J., Ma, Y., Zhang, J., & Zhang, Y. (2022).

Motor imagery EEG decoding using manifold embedded

transfer learning. Journal of Neuroscience Methods, 370,

109489.

Khademi, Z., Ebrahimi, F., & Kordy, H. M. (2022). A

transfer learning-based CNN and LSTM hybrid deep

learning model to classify motor imagery EEG signals.

Computers in biology and medicine, 143, 105288

ICT4AWE 2024 - 10th International Conference on Information and Communication Technologies for Ageing Well and e-Health

220