3652 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL.

31, 2023

Multi-Channel FES Gait Rehabilitation

Assistance System Based on
Adaptive sEMG Modulation
Chunfu Lu , Ruite Ge , Zhichuan Tang , Xiaoyun Fu , Lekai Zhang , Keshuai Yang , and Xuan Xu

Abstract— Functional electrical stimulation (FES) can be Index Terms— BILSTM, EMG prediction, functional
used to stimulate the lower-limb muscles to provide walk- electrical stimulation, gait rehabilitation.
ing assistance to stroke patients. However, the existing
surface electromyography (sEMG)-based FES control meth-
ods mostly only consider a single muscle with a fixed I. I NTRODUCTION
stimulation intensity and frequency. This study proposes
TROKE has become the second leading cause of death
a multi-channel FES gait rehabilitation assistance system
based on adaptive myoelectric modulation. The proposed
system collects sEMG of the vastus lateralis muscle on
S globally. It affects roughly 13.7 million people and kills
around 5.5 million annually [1]. Therefore, how to accel-
the non-affected side to predict the sEMG values of four erate recovery from stroke and restore walking ability has
targeted lower-limb muscles on the affected side using
a bidirectional long short-term memory (BILSTM) model. become a research focus in stroke treatment [2]. Many stroke
Next, the proposed system modulates the real-time FES patients suffer from foot pronation, foot drop, and lower
output frequency for four targeted muscles based on the limb spasm due to the reduced lower-limb muscle strength,
predicted sEMG values to provide muscle force compensa- and asymmetrical gaits can often be observed when they
tion. Fifteen healthy subjects were recruited to participate
are walking [39]. Therefore, improving asymmetrical gaits is
in an offline model-building experiment conducted to eval-
uate the feasibility of the proposed BILSTM model in a major focus of rehabilitation and patients’ top goal [36].
predicting the sEMG values. The experimental results Functional electrical stimulation (FES) has been widely used
showed that the R2 value of the best-obtained prediction to support walking and rehabilitation. The main principle of
result reached 0.85 using the BILSTM model, which was this walking-assistance stimulation is that muscles contract
significantly higher than that using traditional prediction and relax rapidly when subjected to electrical stimulation, thus
methods. Moreover, two patients after stroke were recruited
in the online assisted-walking experiment to verify the producing muscle contraction. The main function of electrical
effectiveness of the proposed walking-assistance system. stimulation is to complete the movement of limbs through
The experimental results showed that the activation of the muscle contraction [35]. The FES has been mostly used to
target muscles of the patients was higher after FES, and support rehabilitation after stroke [3], [4]. The muscle con-
the gait movement data were significantly different before
traction induced by the FES can trigger movement awareness
and after FES. The proposed system can be effectively
applied to walking assistance for stroke patients, and the and promote muscle recovery [5], [6].
experimental results can provide new ideas and methods In addition to the application in the rehabilitation pro-
for sEMG-controlled FES rehabilitation applications. cess, FES has been used to correct the hemiparetic gait
and assisted patients walking in real time. Stein et al. [7]
Manuscript received 29 December 2022; revised 9 June 2023
and 31 August 2023; accepted 6 September 2023. Date of publication proposed the WalkAide (WA) footdrop stimulator that uses
11 September 2023; date of current version 18 September 2023. This the FES to treat foot drop and achieved good efficacy and
work was supported in part by the Key Research and Development acceptance. By detecting and performing the FES according
Program of Zhejiang Province under Grant 2022C03148, in part by
the National Social Science Fund of China under Grant 22CTQ016, to a patient’s gait, the patient’s foot drop gait was mitigated.
and in part by the National Natural Science Foundation of China under Watanabe et al. [8] used the FES to help hemiplegic indi-
Grant 72304249. (Corresponding author: Zhichuan Tang.) viduals strengthen voluntary muscle movement during gait
This work involved human subjects or animals in its research. Approval
of all ethical and experimental procedures and protocols was granted by rehabilitation. In most of the aforementioned studies, only a
the Ethics Committee of the Zhejiang University of Technology. single muscle was electrically stimulated. However, patients
Chunfu Lu, Ruite Ge, Xiaoyun Fu, Lekai Zhang, Keshuai Yang, and with gait problems often have difficulties with more than
Xuan Xu are with the Industrial Design Institute, Zhejiang University of
Technology, Hangzhou 310023, China. one muscle [37]. Namely, human gait is maintained by the
Zhichuan Tang is with the Industrial Design Institute, Zhejiang Univer- coordinated activities of different muscles in the hip, knee, and
sity of Technology, Hangzhou 310023, China, and also with the Faculty ankle joints [9]. Therefore, regular gait changes represent the
of Science and Technology, Bournemouth University, BH12 5BB Poole,
U.K. (e-mail: ztang@zjut.edu.cn). result of the joint action of several muscles. When the human
Digital Object Identifier 10.1109/TNSRE.2023.3313617 body walks, all parts of the body are in motion, and there is

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
LU et al.: MULTI-CHANNEL FES GAIT REHABILITATION ASSISTANCE SYSTEM 3653

a periodic relationship between the activities of all muscles. LSTM network and a multi-layer perceptron (MLP) algorithm
However, there has been little research on gait assistance for to classify human gaits. They developed a low-cost gait phase
the combined electrical stimulation of multiple muscles based detection system that can assess the gait cycle in real time
on the laws of human movement. based on surface electromyographic data. However, the LSTM
In the existing research on walking assistance, the elec- network has the limitation that the output tends to converge
trical stimulation has been mostly of fixed intensity and based on a rectilinear pattern because the inputs are used in
frequency [10]. These two parameter values and timing of chronological order. To address this problem, the bidirectional
electrical stimulation must be strictly controlled for each long short-term memory network (BILSTM) adds reverse
muscle when it is applied to the human leg muscles. This learning to the original forward learning. Previous studies have
is because an inappropriate stimulation parameter values or shown that the BILSTM can capture additional features related
timing might cause gait disorders in patients [11]. Therefore, to the time series data for achieving more complex predictions
it is challenging to achieve a symmetrical gait. Thus, how to over long sequences. However, the traditional LSTM model
assess the timing and parameter values of electrical stimulation cannot capture these features [18]. Therefore, in this work, the
has been an important issue in areas of stroke rehabilitation BILSTM is used to predict the EMG signals of the affected
research. The adaptive electrical stimulation must be modu- lower limb muscles.
lated according to the real-time gait information so that the This paper proposes a multi-channel FES gait rehabilitation
muscle force can be dynamically compensated. Accordingly, assistance system based on the contralateral EMG signal
an accurate prediction of gait information of the affected limb control. First, the EMG signals of the vastus lateralis muscle
is necessary. Some studies have predicted gait information are recorded in real time using surface EMG acquisition
of the affected side based on the electromyography (EMG) equipment. Next, the EMG values of the four muscles (rectus
signal and motion data of the non-affected side and con- femoris muscle, biceps femoris muscle, tibialis anterior mus-
trolled the intensity of electrical stimulation to optimize the cle, peroneus brevis muscle) are predicted by the BILSTM
walking-assistance performance. Chen et al. [11] proposed a network prediction model. Using a four-channel electrical
gait prediction method for the affected side based on a patient’s stimulation device, the appropriate intensity and frequency
EMG signal during walking and adaptively modulated the of electrical stimulation compensation is determined for a
electrical stimulation parameters. They have proved that the particular patient to assist in walking.
electrical stimulation curve modulated according to the EMG The main innovations in this work are as follows:
signal can activate patients’ muscles and achieve a better effect (1) Electrical stimulation of multiple muscles is used to
than the original trapezoidal stimulation. Jiang et al. [12] achieve real-time walking assistance for hemiplegic
developed a walking support system that adaptively adjusts stroke patients.
the electrical stimulation intensity of the affected lower limb (2) Variable EMG signal is predicted by the BILSTM net-
according to the kinematic data of the non-affected side, and work, and the electrical stimulation power is adaptively
the achieved effect of walking support was better than that modulated according to the EMG signal to realize walk-
of the fixed electrical stimulation. The relationship model ing assistance.
between healthy and affected limbs established based on the (3) Compared with the existing electrical stimulation sys-
EMG and motion data can be effectively used for adaptive tems, our multi-channel FES gait rehabilitation assis-
electrical stimulations for walking support. In the related liter- tance system uses fewer input channels to control the
ature, there have been fewer studies that used the EMG data of real-time output of more channels.
the non-affected side to predict the EMG value of the affected The rest of the paper is organized as follows. Section II
side. However, there is a certain correlation between the EMG describes the principle framework and development process
data and muscle strength, which can be used to characterize of the electrical stimulation system. Section III introduces
the activation of muscles of the affected limb [13]. When the experiment, including the experimental personnel, process,
the difference in the muscle force (myoelectric difference) equipment, and data processing. In Section IV, the experi-
between healthy and affected limbs is determined, the FES mental results are obtained and expounded. In Section V, the
can be modulated with an appropriate intensity according to experimental results are analyzed and discussed. Section VI
the obtained difference in magnitude of the two myoelectric draws the conclusions.
signals.
Previous studies have used the EMG signals to predict a gait II. S YSTEM D ESIGN
pattern employing different prediction models, such as BP neu- This gait rehabilitation system with electrical stimulation
ral network (BPNN) and support vector machine (SVR) [14], consists of a signal transmission and processing subsystem
[15], [16]. Although the achieved results are acceptable, and an electrical stimulation output subsystem. The signal
human gait is a periodic movement, so the EMG data have transmission and processing subsystem includes an EMG
time-series characteristics. As well-known, a long short-term acquisition module and a model training module, which are
memory network (LSTM) network can capture the time series mainly responsible for EMG data acquisition and processing
characteristics of gait data well and thus has been widely and offline training of a prediction model. The electrical
applied to the EMG signal-based gait prediction models. stimulation output subsystem includes a control module and
Luo et al. [17] used a surface electromyography signal acqui- an electrical stimulation output module, whose functions are
sition system and a plantar pressure sensor combined with an mainly to manage the electrical stimulation output and provide
3654 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

Fig. 1. System framework.

walking assistance to patients. In the entire application process

of the multi-channel FES gait rehabilitation assistance system,
the EMG acquisition module collects the EMG data of the
vastus lateralis muscle and transmits it to a computer through
a Bluetooth module. Then, Matlab 2021a software is used to
predict the EMG value and process the difference between the
module’s muscle EMG data and the predicted EMG value by
a neural network. Afterward, the obtained difference value is
transmitted to an Arduino through a Bluetooth module, which
controls the STM32 development board. Finally, the STM32 Fig. 2. Four EMG prediction models. The myoelectricity of the vastus
development board provides the electrical stimulation output lateralis muscle on the non-affected side is matched with that of the four
muscles for training, and four prediction models are designed.
to achieve the function of real-time electrical stimulation to
the main walking muscles of the human body. The specific
system framework is shown in Fig. 1.
by superimposing and averaging 30 walking cycles as a
A. Signal Transmission and Processing Subsystem healthy standard for the EMG data. This healthy standard
1) EMG Acquisition Module: The EMG acquisition module is used to calculate the missing value of the EMG data of
represents a combination of an EMG sensor module (Wuxi the affected side in real time. Moreover, the average rectified
Sichrui Co., Ltd., Wuxi, Jiangsu, China) and a Bluetooth EMG (AEMG) value of the vastus lateralis muscle and the
module (HC-05). The collected EMG data from the lateral AEMG value of the rectus femoris (biceps femoris, anterior
thigh muscle are transmitted to the computer via Bluetooth tibialis, and peroneus brevis) are used in neural network
and used as input for the prediction model. training. Finally, four EMG prediction models of the four
2) Model Training Module: It is necessary to collect the EMG contralateral muscles are designed, as shown in Fig. 2. When a
data of the vastus lateralis muscle on the non-affected side and patient uses the rehabilitation system, the electrical stimulation
the four muscles on the affected side (rectus femoris, biceps collection system collects the EMG data of the vastus lateralis
femoris, tibialis anterior, and peroneus brevis) for the BILSTM muscle on the non-affected side and uses it as an input of the
network training. Function of the rectus femoris and biceps rehabilitation system. Then, the trained neural network model
femoris are closely related to thigh swing and knee extension is employed to predict the AEMG value of a target muscle and
in stroke patients, while the weakness of the tibialis anterior perform a BILSTM regression prediction. Finally, the STM32
and peroneus brevis represents the main factor affecting the development board is used to obtain the electrical stimulation
foot landing symptoms in stroke patients [19], [20]. Since the output of the required parameter values in each channel in real
vastus lateralis of the thigh shows good muscle activity and time.
significant EMG activity during the entire gait cycle and it also During the EMG data acquisition process, a large number
exhibits excellent periodicity [14], it can be used as an input of long time series are generated using a high-frequency EMG
signal to determine the timing and parameter values of electri- collector. Therefore, in this study, the BILSTM algorithm is
cal stimulation output. In addition, it is needed to collect the used to process and predict the EMG data. The BILSTM
EMG data of the four muscles on the non-affected side (rectus algorithm combines reverse learning and the original LSTM
femoris, biceps femoris, tibialis anterior, and peroneus brevis) network. The same input sequence is connected to the forward
and use the EMG signal’s waveform in a single cycle generated and backward LSTMS, and the hidden layers of the two
LU et al.: MULTI-CHANNEL FES GAIT REHABILITATION ASSISTANCE SYSTEM 3655

Fig. 5. Electronic development board for electrical stimulation output.

Fig. 3. The BILSTM structure.

is used to receive the information on a difference between the
EMG signal predicted by the neural network on the computer
and the EMG healthy standard and forward it to the Arduino
board. The Arduino board further forwards the received data to
the electrical stimulation output module to control the output
value of the four channels.
2) Electrical Stimulation Output Module: A development
board for an electrical stimulation output system is designed,
as shown in Fig. 5. Compared with the existing electrical stim-
ulation output systems, the proposed system has four output
channels and can independently obtain electrical stimulation
output of different intensities and frequencies. The circuit
board uses an STM32 as a controller, adopts the RS-232
interface, and can communicate with other development boards
to realize the electrical stimulation output control.
Fig. 4. The internal structure of a single LSTM unit. The computer In the developed system, the STM32F103RCT6 is used
performs self-learning on the cell state at the previous moment, and the as an MPU, the external 8-MHz crystal oscillator is used
predicted output value of the myoelectricity at the previous moment and
the input myoelectricity value of the current moment are used to obtain to achieve the main frequency, and the GPIO is controlled
vectors z, zi, zf, and zo after performing the combined transformation. to provide the output to the 74HC541 line driver, which is
used to control the optocouplers and isolate the front and
rear circuits. The output signal is driven and amplified by an
networks are connected both mutually and to the output OTL circuit to generate periodic square waves with different
layer to realize prediction. At the same time, the connection frequencies to realize electrical stimulation. The 12-V signal
structure of forward and backward time series are learned, is fed to the DC5 connector, decompressed to 5 V by a
as shown in Fig. 3. The internal structure of a single LSTM DC-DC chip, and then processed by the LDO chip to pro-
unit is shown in Fig. 4, where X t is the feature vector of duce stable 3.3-V voltage for the STM32. In addition, the
the time series in the LSTM network. Through the function 12-V DC power supply is connected to the capacitor via
transformation of Xt , which can be realized by a computer, a high-frequency transformer, and the OB2263 control field
four new vectors z, z i , z f , and z o are obtained and used as effect tube is disconnected to isolate the voltage boost. Finally,
four inputs to the LSTM network; z i , z f , and z o are input the secondary signal is rectified and filtered to maintain a
vectors of the input, forget, and the output gates, respectively. continuous power supply. When the supply voltage reaches
The four inputs (i.e., z, z i , z f , and z o ) are input to the network the target value, the reference voltage is intercepted by the
simultaneously and processed by the activation function, for voltage over resistors, and the optocoupler is driven by the
instance, a sigmoid function. Further, c and c′ in the network LM431 to send negative feedback to the OB2263 control field
represent the current and updated states of the memory cells, effect tube, which turns off the MOS to control the voltage. For
respectively. After the updating process of memory cells by power supply, the proposed design uses a 2,000-MAh lithium
the activation functions, the relationship between the input val- battery, and a Type-C connector is used to charge and transfer
ues z, z i , and z f can be expressed by c′ = g(z)f(z i ) + cf(z f ). programs.
The final output is expressed as h t = h(c′ ) f(z o ); vector h(c′ ) The electrical stimulation can be adjusted in three main
is obtained based on the information in the memory cell by the ways, namely by adjusting the amplitude, width, or frequency
activation function h, and vector f(z o ) is formed by applying of a pulse signal. Increasing the amplitude and width of a
activation function f to vector zo . The output ht denotes the pulse can affect the intensity of a single electrical stimu-
AEMG of the target muscle. lation, whereas increasing the pulse frequency can amplify
muscle contraction if the pulses arrive rapidly enough to sum-
B. Electrical Stimulation Output Subsystem mate, finally achieving a stronger electrical stimulation effect,
1) Control Module: The control module consists of a Blue- as shown in Fig. 6. The proposed design uses a bidirectional
tooth module and an Arduino board. The Bluetooth module electrical stimulation pulse with a pulse amplitude of 50 ma,
3656 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

Fig. 6. (a) Changes in muscle force after a single electrical stimulation;

(b) the superimposed effect of multiple high-frequency stimulations on
muscle force.
Fig. 8. Experimental pictures of subjects participating in the experi-
ments: (a) healthy male person; (b) male patient who recovered from
a stroke.

Considering that there could be great differences in the stride

length of the subjects, the height of the subjects was limited
to ensure that each participant could walk on the treadmill
according to the predetermined step frequency while main-
taining the same running speed. The height requirements for
Fig. 7. Illustration of the electrical stimulation compensation.
female and male subjects were between 160 cm and 165 cm
and between 170 cm and 175 cm, respectively.
In addition, two patients who recovered from a stroke
a pulse width of 200 µs, and an adjustable frequency were also invited as representatives to participate in the
of 0–150 Hz. According to the previous relevant research on experiments to verify the reliability of the proposed system
electrical stimulation, the usual pulse amplitude is set between further (as shown in Fig. 8b). Both stroke patients were
zero and 130 mA. The pulse amplitude of 50 mA can be effec- male. Patient 1 was a 61-year-old male with left hemiplegia
tive in most patients without causing much discomfort [21]. caused by a stroke and had an unsteady gait with obvious
The pulse frequency is adjusted between 0-150 Hz. Since compensatory gait. His reduced thigh muscle strength hindered
the main objective of this study is to help patients to walk, adequate thigh lift and complete gait, leading to tremors during
it is primarily necessary to strengthen the muscle strength of walking. Furthermore, he demonstrated inadequate calf muscle
the affected side of patients. Namely, adjusting the frequency strength and mild foot drop. In addition, his calf muscle
of electrical stimulation can effectively help patients to gain strength was insufficient, and the ankle showed a slight foot
muscle force in a short time [23], [24]. Although using a drop phenomenon. Patient 2 was 58 years old and had left
too-high frequency can accelerate muscle fatigue [22], it has hemiplegia caused by a stroke and a slight foot drop on the left
been documented that increasing the frequency of electrical side, and his overall gait was more stable than that of patient 1.
stimulation contributes to muscle force more than increas- The experimental procedure of this study was approved by
ing the pulse amplitude and pulse width [24]. Therefore, the Ethics Committee of the Zhejiang University of Technol-
in this study, the frequency is changed to provide electrical ogy, and all subjects and patients provided signed informed
stimulation to patients to improve their muscle force output. consent forms.
The difference between the EMG healthy standard value of
each muscle on the non-affected side and the corresponding B. Experiment Procedure
EMG value of each muscle of the affected leg predicted Two experiments were conducted to verify the effective-
by the neural network is determined in each gait cycle. ness of the proposed system. The reliability of the proposed
The resulting difference curve is compensated by electrical algorithm was tested by an offline model-building experiment.
stimulation. Finally, the numerical range of the difference is The signals of the five leg muscles of 15 healthy volunteers and
mapped to the electrical stimulation frequency value range and two patients with hemiplegia were measured, and the BILSTM
used to control the final electrical stimulation frequency for algorithm was used to predict the EMG data of the four
each channel. The electrical stimulation support principle is major walking muscles of the affected side based on the input
illustrated in Fig. 7. EMG data of the vastus lateralis of the non-affected side.
After the reliability of the proposed algorithm was proven, the
III. E XPERIMENTAL R ESULTS online assisted-walking experiment was conducted to verify
A. Information on Subjects Participating in Experiments the corrective effect of the proposed electrical stimulation
Fifteen healthy subjects were selected to verify the reliabil- rehabilitation assistance system on the patients’ gait and its
ity of the proposed prediction algorithm (as shown in Fig. 8a). effect in providing walking support.
The speed and step frequency values were limited by using 1) Offline Model-Building Experiment: The EMG data of the
a treadmill and a metronome respectively, approximately legs were collected when the subjects were walking using a
at 70 steps per minute with a treadmill speed of 60m/min [38]. metronome and treadmill and controlling the pace at about
LU et al.: MULTI-CHANNEL FES GAIT REHABILITATION ASSISTANCE SYSTEM 3657

70 steps per minute [25]. The collected data were used to similar; the excitation threshold of the electrical stimulation
train the neural network. The discrepancy degree between of the thigh was 50 Hz, and the patient felt mild discomfort
the EMG value predicted by the proposed algorithm and the at 100 Hz. The excitation threshold of the calf was 50 Hz,
actual measured healthy-standard EMG value was measured. and mild discomfort occurred at 200 Hz. On this basis, the
As mentioned in previous studies, an unstable step frequency frequency ranges of electrical stimulation were determined for
of subjects at a controlled speed can cause large differences in the four channels. The threshold of electric stimulation for the
the muscle stretch length, resulting in a decrease in the EMG muscles in the patient’s legs is 50Hz. Therefore, we set the
data quality [26]. Therefore, in this experiment, the patients’ electric stimulation occurrence plate to receive a predictive
speed and step frequency were controlled. data signal every 0.02 seconds to adjust the output frequency
In this experiment, the subjects’ speed on a treadmill and of the electrical stimulation in real-time. We have controlled
step frequency were controlled by a metronome. The subjects the single run time of the entire system within 0.02 seconds
wore an EMG sensor (MP150) when walking on the treadmill, to ensure the real-time requirement of the system operation.
and the EMG data of the five muscles (i.e., rectus femoris, The patients performed a walking test while wearing the
biceps femoris, tibialis anterior, peroneus brevis, and vastus electrical stimulation system. The walking test was performed
lateralis) were recorded. Under the condition of wearing an on a treadmill. According to the related research, the walking
EMG sensor, subjects walked for five minutes on a treadmill speed used in the experiment belonged to the slow walking
at a constant step frequency according to the beeps of a conditions. If the frequency of slow walking was not limited,
metronome. [40] The controlled walking speed was 70 steps it would be difficult for muscles to maintain the condition of
per minute, and four trials were performed, with a five-minute the same length [28]. The EMG data quality decreased to a
muscle relaxation between adjacent trials to avoid fatigue in certain extent when only the speed was controlled but not the
patients. Timing started when the subject made the first steady step frequency. Since the step frequency in the experiment
step on the treadmill. was controlled by a metronome, there were no significant
During the BILSTM training, the EMG data of each subject differences in the variations of muscle contraction frequency
was used to train the model. To test the reliability of the and muscle length when walking within each cycle. The
BILSTM for predicting the surface EMG of different mus- walking test was divided into two phases. In the first phase,
cles, the EMG signals of 15 subjects and two patients were patients were first asked to walk for two minutes without
measured and compared with the corresponding predicted receiving any electrical stimulations. A total of four trials were
values. There were certain differences in the surface EMG for performed. Then, a two-minute break was provided between
different subjects, so the BILSTM needed to be readjusted to trials to avoid muscle fatigue. In the second phase, patients
achieve a satisfactory prediction effect for different subjects. were subjected to electrical stimulation while walking. Patients
In this study, the BILSTM model was evaluated by the were asked to walk for two minutes, and a total of four trials
tenfold cross-validation method. The dataset was split into were performed. Similarly, a five-minute break was ensured
two parts: 80% of the data were used for training data, between adjacent trials to allow the muscle to relax fully after
and the remaining data were used as test data. We set up the electrical stimulation.
a sliding window with an input window size of 100ms and We also conducted a separate data collection of muscle
an output window size of 20ms for our prediction model. activation. The patients were asked to lift the thigh as much
This minimized the computational load of the model during as possible and actively contract the calf during the late thigh
online assisted-walking experiment and ensured its real-time swing phase of walking. In this way, the EMG data of a
performance. The model’s inputs were the mean rectified EMG single contraction and large voluntary contraction of the rectus
values of the vastus lateralis muscle from the previous 100ms, femoris and biceps femoris were measured. During the swing
starting from the current time point, and its outputs were the phase of walking, the patients were asked to decrease the
mean rectified EMG values s of the four target muscles for angle between the calf and ankle as much as possible while
the next 20ms. The proposed algorithm was compared with the maintaining a continuous contraction of the tibialis anterior
AEMG+BP and AEMG+SVR algorithms. The AEMG is an and peroneus brevis muscles. The EMG data of a single
important feature for characterizing changes in the EMG data contraction and large voluntary contraction of tibialis anterior
in the time domain, and it was used in the regression analysis muscle and peroneus brevis muscle were measured. The above
in the studies by J. R. Potvin and L.R. Bent et al. [27]. The actions need to be performed only once in a single gait cycle,
mean absolute error percentage (MAPE), root mean square and ten gait cycles were measured. The measurements were
error (RMSE), and R2 were used to evaluate the fitting degree performed according to the above requirements both with and
of the EMG value predicted by the BILSTM. without the electrical stimulation. Based on the collected data,
2) Online Assisted-Walking Experiment: After obtaining the the muscle activation degree was calculated. We analyzed the
patient’s EMG data prediction model and programming the mean values of the signals collected during 10 gait cycles
gait rehabilitation assistance system, the electrical stimulation under two walking conditions, and integrated them into muscle
excitation threshold and the limit value of each muscle in the activation calculations.
two patients were tested. The electrical stimulation system
used bidirectional square-wave pulses with a pulse width C. Experimental Equipment and Data Processing
of 200 µs and a variable frequency at an electrical stimulation 1) Experimental Equipment: In both experiments, the EMG
intensity of 50 mA. The test results of the two patients were acquisition equipment included an mp150 EMG acquisition
3658 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

indicators included the swing angle and acceleration of the

knee and ankle joints before and after electrical stimulation.
These two indicators could be used to describe the overall
condition of a gait. If there was no significant difference in
the EMG data, the independent flexion angle of the joint and
acceleration perpendicular to the motion direction should be
investigated to judge whether there was a significant difference
in gait before and after the electrical stimulation to illustrate
the effectiveness of real-time electrical stimulation in assisting
gait.
In addition to the kinematic data, the difference between
the EMG data when the patients walked voluntarily and the
EMG data when they walked wearing the electrical stim-
Fig. 9. (a) Muscle position in the EMG acquisition process; (b) EMG ulation system was determined. In addition, the degree of
acquisition equipment; (c) motion sensors.
muscle activation with both sets of EMG data was measured.
Since the proposed electrical stimulation assistance system
was primarily designed to help patients walk better, the data
of a single voluntary movement of the patients during the
experiment and data of the maximum voluntary contraction
under the electrical stimulation were used to calculate the
muscle activation degree. Because the artifacts produced by the
electrical stimulation interfere with the calculation of muscle
activation degree, the following methods were used to remove
the artifacts of electrical stimulation:
Fig. 10. (a) VL raw EMG data while walking; (b) VL aEMG. (a) High-pass filtering was performed on the original EMG
using a Butterworth high-pass filter, with an order of
four and a cutoff frequency of 15 Hz;
system from a biopac physiological recorder (USA) and an (b) The algorithm of a comb filter defined by (1) was used
EMG sensor module (Wuxi Sichrui Co., Ltd.); the acquisition for filtering, and the spontaneous and induced EMG
frequency was 1 kHz in each case. parts of EMG were retained.
Since the proposed system focused on gait assistance x (n) − x(n − N Stim )
functions, three nine-axis accelerometers (Witt Electronic y (n) = √ (1)
2
Intelligent Technology Co., Ltd., China) was used in the
In (1), x(n) is the raw EMG signal at a sampling time n;
trial phase of the online assisted walking to analyze relevant
Nstim is the inter-stimulus time expressed in the number
kinematic data of a patient. The nine-axis transducers a, b, and
of samples, and y (n) is the filtered EMG signal [30].
c were installed in the sagittal plane of the mid-femur, below
the knee joint, and below the ankle joint of the patients. The Next, the muscle activation degree was calculated as
sampling frequency of the nine-axis sensor was 1 kHz. The follows. First, the filtered single conventional motion signal
wearing position of the sensor is shown in Fig. 9. To eval- was divided by the maximum value of the maximum auto-
uate the effect of electrical stimulation on thigh muscles, nomic contraction sEMG signal to obtain the normalized EMG
we assessed the thigh lift angle and the acceleration of the signal e j (t), which was used to solve the nerve activation
knee joint movement, both directly measured by accelerometer intensity u j (t) as follows:
a. To quantify foot drop alleviation, we recorded the swing u j (t) = αe j (t − d) − β1 u j (t − 1) − β2 u j (t − 2) , (2)
angle and acceleration of the ankle joint, and the independent
where α represents the gain coefficient; d is the time delay of
flexion angle of the same joint. While the swing angle and
the electrode; β1 and β2 represent the recursive coefficients;
acceleration were directly captured by accelerometer b, the
α, d, β1, and β2 are constants.
independent flexion angle was determined by integrating data
According to the neural activation degree, u j (t) was used
from accelerometers b and c. Au j (t)
2) Data Processing: A bandpass filter with a passing band to obtain the muscle activation degree a j (t) = e e A −1−1 , where
from 10 Hz to 200 Hz and a 50-Hz notch filter were used A is a nonlinear shape coefficient, representing the nonlinearity
to obtain the filtered EMG data [29], and the AEMG were relationship between the nerve activation intensity u j (t) and
extracted manually, a 300-ms average rectified window func- the muscle activation intensity a j (t); A was set to −1.5 in
tion was used for smoothing. The AEMG was used as the this study.
input data of the BILSTM. The EMG signal after processing
is shown in Fig. 10. IV. R ESULTS
The change in kinematic data before and after the applica- A. Offline Model-Building Experimental Results
tion of the electrical stimulation system was used to evaluate The results of the offline model-building experiment showed
the effectiveness of the proposed system. The main verification that due to the complexity of the surface EMG signal, there
LU et al.: MULTI-CHANNEL FES GAIT REHABILITATION ASSISTANCE SYSTEM 3659

Fig. 11. The EMG prediction results of a single gait cycle of the main walking muscles of a representative subject under different prediction
algorithms.

TABLE I TABLE II
A LGORITHM ACCURACY C OMPARISON R ESULTS P REDICTION ACCURACY R ESULTS OF THE BLISTM ON THE
H EALTHY AND A FFECTED S IDES OF THE PATIENTS

by the two classical algorithms, but the results’ accuracy

was poor. Compared with the two classical algorithms, the
fitting effect of the proposed BILSTM network on the EMG
values of the four muscles was much better. Regarding the
R2 value, the lowest R2 of the BILSTM model reached
were many small waveforms with lower peak amplitude in 0.62, and R2 of the BILSTM model for the EMG of the
addition to the main EMG signal waveform. The proposed tibialis anterior muscle (TA), which could not be predicted
BILSTM could predict the periodicity of the main waveform, by the two classical algorithms, was 0.83. The R2 value of
but it could not accurately predict the smaller waveforms. the best-obtained prediction result reached a value of 0.85.
However, the prediction curve of the main EMG signal essen- Based on the prediction results, it was difficult to predict small
tially passed through the vertex-connection lines of the smaller waveforms by the BILSTM network accurately, which affected
waveform peaks. The results indicated that the BILSTM based the fitting results to a certain extent. Thus, when the small
on the surface EMG of the vastus lateralis could predict the waveforms were ignored, and only the main waveform was
EMG of the four leg muscles. The prediction effect of the considered, the R2 value reached 0.8 or more, showing that
proposed BILSTM model is shown in Fig. 11. the model could meet the requirement of prediction [31]. The
The results of different algorithms showed that the proposed comparison results of the three algorithms showed that the
BILSTM model was far superior to the traditional SVM and proposed BILSTM network had a strong ability to predict and
BP models in predicting complex EMG signals. The numerical fit complex EMG signals.
results of the algorithms’ comparison are shown in Table I. In the follow-up experiment, the patients performed an
The MAPE, RMSE, and R2 of the EMG regression prediction online assisted-walking experiment at a pace of 70 steps. The
results for 15 subjects were compared. Regarding the MAPE R2 values of the predicted data of the two patients are shown
and RMSE, the proposed BILSTM network performed better in Table II. The EMG predicted results of the patients were
than the two classical algorithms. As for R2 of the model fitting not significantly different from those of the healthy subjects.
effect, the prediction results of the two classical regression
algorithms for the tibialis anterior muscle (TA) had R2 values
less than or equal to zero. This indicated that the two classical B. Online Assisted-Walking Experimental Results
algorithms could not predict the EMG data of the tibialis We analyzed the gait kinematic data of patients using sta-
anterior muscle (TA) in this experiment. The surface EMG tistical parametric mapping (SPM) method, one-dimensional
values of the remaining three muscles could be predicted SPM (1D-SPM) paired t-test was performed to determine
3660 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

Fig. 12. (a) Thigh lift angle data; (b) SPM metric between thigh lift angle
data before and after the FES; (c) knee joint acceleration data; (d) SPM
metric between knee joint acceleration data before and after the FES.
Fig. 14. (a) Ankle joint swing angle data; (b) SPM metric between
ankle joint swing angle data before and after the FES; (c) ankle joint
acceleration data; (d) SPM metric between ankle joint acceleration data
before and after the FES.

Fig. 13. (a) Thigh lift angle data of the left leg after the FES application
and right leg; (b) SPM metric between thigh lift angle data of the left leg
after the FES application and right leg.
Fig. 15. (a) Ankle-independent flexion angle data; (b) SPM metric
between ankle-independent flexion angle data before and after the FES.

differences in patients’ gait data before and after the appli-

cation of electrical stimulation. We used open-source code
(www.spm1d.org, accessed on 7 April 2021) in MATLAB.
The significance level was set at α = 0.05.
The comparison of the patient’s thigh lift angle before
and after using the FES system showed that the value of
the thigh lift angle increased and the variance of the swing
angle decreased when the patient walked with the FES system,
Fig. 16. (a) Acceleration perpendicular to the ankle joint; (b) SPM metric
as shown in Fig. 12(a). Moreover, the forward acceleration of between acceleration perpendicular to the ankle joint data before and
the patient’s knee joint swing was analyzed, and it was found after the FES.
that the forward acceleration was higher with FES, as shown
in Fig. 12(c), which was due to the larger change in swing
angle during each walking cycle. The 1D-SPM paired t-test of showed that there was no significant difference in ankle swing
the data found a significant difference in thigh lift angle before angle and forward acceleration before and after the FES
and after using FES (p < 0.05). The thigh lift angle before and application (p > 0.05), as shown in Fig. 14.
after using FES exceeded the threshold of significance during Since the ankle swing angle was not significantly different
28% to 48% and 54% to 77% of the gait cycle There was before and after the FES application, a separate measurement
no significant difference in forward acceleration before and of the ankle flexion angle was performed, as shown in Fig. 15.
after the FES application. However, the acceleration differed The ankle flexion angle data for 30 cycles were obtained, and
significantly between I before and after using FES in the range an 1D-SPM paired t-test was performed. The results showed
from 70% to 90% of the gait cycle (Figure 4; p < 0.05). that the ankle flexion angle was larger after the FES than
The data of the thigh lift angle of the left and right before it, and there was a significant difference (p < 0.05).
leg were obtained by experimental measurements, as shown The data exceeded the threshold of significance during 40 to
in Fig. 13, and an 1D-SPM paired t-test was performed. The 70 percent of the gait cycle.
results indicated no significant difference in the thigh lift angle The compensatory gait present in the patient’s ankle joint
between the left leg after the FES and right leg (p > 0.05). represents a pronounced circular gait. The calf and ankle joints
The data of the thigh lift angle of the left after using FES and arch to the outside of the body and then swing to the front
right leg only exceeded the threshold of significance during of the body. Therefore, the acceleration perpendicular to the
35 to 47 percent of the gait cycle. The results also showed ankle joint was measured, as shown in Fig. 16.
that the kinematic values of the limbs on the affected side After stacking and averaging multiple datasets, a 1D-SPM
gradually approached those on the non-affected side after the paired t-test was conducted. There was no statistically signifi-
FES application. cant difference in the data regarding acceleration perpendicular
Next, an 1D-SPM paired t-test was performed for ankle to the ankle joint before and after the use of FES (p > 0.05).
swing angle and ankle forward acceleration, and the results However, compared to the two groups of gait data, the variance
LU et al.: MULTI-CHANNEL FES GAIT REHABILITATION ASSISTANCE SYSTEM 3661

small and its position was deep, so the quality of the measured
EMG data was lower than that of the other three muscles,
leading to poor prediction results. At the same time, there was
certain instability in patients in the gait cycle. Although certain
abnormal data were excluded during the model training, the
overall quality of the EMG data had a certain influence on
the prediction results. Gabel et al. [32] have also concluded
that the smoother and less fluctuating the EMG signal is
Fig. 17. Muscle activation data in the major walking muscles after processing, the better the fitting effect of the prediction
(a) tibialis anterior muscle activation; (b) peroneus brevis muscle acti- algorithm will be. In addition, the experimental results showed
vation; (c) rectus femoris muscle activation; (d) biceps femoris muscle that different algorithms had different prediction effects on the
activation.
muscle EMG; the proposed BILSTM performed significantly
better than the BP neural network and SVM regression. This
in the acceleration in the direction perpendicular to the ankle was because the walking EMG data were long-time-series
was significantly reduced after the FES. This indicated that data, and its time-series features could effectively help in the
after the FES application, the patient’s gait was more stable in model construction process. The BILSTM could extract the
the direction perpendicular to the ankle than before the FES. time-series features and improve model accuracy. Lu et al. [33]
In addition, muscle activation was used to compare changes have drawn the same conclusion in their study on joint angle
in the EMG data before and after the use of the FES system. prediction based on the surface EMG.
The calculated muscle activation data are shown in Fig. 17. The experimental results showed that the proposed FES sys-
According to the muscle activation results, the total voluntary tem could effectively assist patients in walking. The kinematic
activation of the two calf muscles was lower without the FES. data demonstrated that the proposed electrical stimulation
The muscle activation of the tibialis anterior muscle at the support system could effectively help patients to improve
maximum muscle activity was only 14.3% of that after the the swing amplitude and flexion angles of the knee and
electrical stimulation. This indicated that after electrical stim- ankle joints. Therefore, the overall gait of the leg gradually
ulation, the muscle activation of the tibialis anterior muscle approached the gait of the healthy leg after using our FES
increased to approximately 600% of the original value. The system. Watanabe et al. [8] found that low-frequency electri-
muscle activation of the peroneus brevis muscle was better cal stimulation combined with the human gait could effectively
than that of the tibialis anterior muscle, and its muscle acti- help the foot dorsiflexion of stroke patients to approach a
vation without the electrical stimulation was 22.7% of that normal gait, which is similar to the results presented in this
after the electrical stimulation. The rectus femoris had local study. At the same time, when the stimulus was used to assist
peaks in the support phase of the entire thigh gait and in the gait, the sudden lateral acceleration signal under the abnormal
lift phase of the thigh. In the support phase of the thigh, the gait was removed in some cycles. The results suggested
activation level of the rectus femoris muscle, when the EMG that the electrical stimulation increased the muscle force and
value reached its peak, was 32.7% of that after the electrical standardized the sequence of muscle force application; thus,
stimulation. In the thigh swing and lift phase, the muscle the use of electrical stimulation may be an effective inter-
activation of the rectus femoris muscle without the electrical vention strategy for correcting compensatory gait patterns in
stimulation was 42.7% of that after the electrical stimulation. patients. The muscle activation measurement results indicated
The biceps femoris mainly controlled the knee flexion during that the residual motor function of the thigh was better than
the thigh swing phase, and the muscle activation of the biceps that of the calf in terms of the EMG value. After electrical
femoris without the electrical stimulation was 35.5% of the stimulation, activation of the rectus femoris and the biceps
peak activation after the electrical stimulation. femoris increased, and the overall gait approached the normal
gait. Further, due to the low residual mobility of the two lower
V. D ISCUSSION leg muscles, the patient could not achieve the required ankle
The feasibility of the proposed algorithm was verified by the flexion angle through autonomic activities without electrical
offline model-building experiment, and the effectiveness of the stimulation, resulting in a slight drop of the patient’s foot.
proposed system was evaluated by the online assisted-walking However, the proposed electrical stimulation system could
experiment. significantly improve muscle activation and help the patient
The experiments on the offline model training showed that to perform ankle flexion effectively.
there were certain differences in the prediction results of As for the hardware of the proposed electrical stimulation
different muscles. The overall prediction effects of the biceps system, due to the advantages of the BILSTM, more output
femoris and tibialis anterior were better than those of the channels can be controlled with fewer inputs. The designed
rectus femoris and peroneus brevis. This could be because the four-channel electrical stimulation development board reduces
biceps femoris and tibialis anterior applied the concentrated the volume of the entire system, so the product can be worn
force throughout the gait cycle, so the value of the main wave- in the form of a belt. Compared with the traditional reha-
form of the surface EMG and the value of the small waveform bilitation devices, the proposed electro-stimulation walking
in other time periods were significantly different, and their aid system is smaller and more portable and thus can be
features were more distinct. The peroneus brevis muscle was used in walking exercises in daily life and is not tied to a
3662 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 31, 2023

specific environment. Marco’s design of wearable electrical [5] S. Okunami, K. Honda, A. Kanada, Y. Nakashima, and
stimulation products with “user-centered approach” shows the M. Yamamoto, “The vibration-induced kinesthetic illusion enhanced
by functional electrical stimulation for neurorehabilitation,” in
convenience of wearable products [34]. Namely, algorithm Proc. World Autom. Congr. (WAC), Aug. 2021, pp. 296–301, doi:
parameters have to be adjusted for particular patients to create 10.23919/WAC50355.2021.9559448.
a neural network model that is adapted to the muscles of that [6] C.-H. Chou et al., “Automated functional electrical stimulation
training system for upper-limb function recovery in poststroke
specific patient. However, for patients, this only means an patients,” Med. Eng. Phys., vol. 84, pp. 174–183, Oct. 2020, doi:
additional phase of the EMG signal acquisition (pre-training 10.1016/j.medengphy.2020.09.001.
phase) before application, and there is no further application [7] R. B. Stein et al., “A multicenter trial of a footdrop stimulator controlled
by a tilt sensor,” Neurorehabilitation Neural Repair, vol. 20, no. 3,
process in the subsequent use by patients. The feasibility pp. 371–379, Sep. 2006, doi: 10.1177/1545968306289292.
of the proposed electrical stimulation system was explored [8] T. Watanabe, S. Endo, K. Murakami, Y. Kumagai, and N. Kuge,
in an experimental setting with two post-stroke participants. “Movement change induced by voluntary effort with low stimulation
Preliminary findings suggest the potential of this system to intensity FES-assisted dorsiflexion: A case study with a hemiplegic
subject,” in Proc. 6th Int. IEEE/EMBS Conf. Neural Eng. (NER), San
assist in gait rehabilitation and promote walking to some Diego, CA, USA, Nov. 2013, pp. 327–330, doi: 10.1109/NER.2013.
extent. However, larger-scale studies are necessary to confirm 6695938.
and extend these initial observations. [9] M. Q. Liu, F. C. Anderson, M. H. Schwartz, and S. L. Delp, “Muscle
contributions to support and progression over a range of walking
The main limitation of this study is that only four muscles speeds,” J. Biomech., vol. 41, no. 15, pp. 3243–3252, Nov. 2008, doi:
are considered in walking assistance, but more lower-limb 10.1016/j.jbiomech.2008.07.031.
muscles are involved in actual walking. In future work, the [10] F. Rossi, P. M. Ros, S. Cecchini, A. Crema, S. Micera, and D. Demarchi,
“An event-driven closed-loop system for real-time FES control,”
output channel number of the proposed electrical stimulation in Proc. 26th IEEE Int. Conf. Electron., Circuits Syst. (ICECS),
system could be increased to achieve more adaptive electrical Genoa, Italy, Nov. 2019, pp. 867–870, doi: 10.1109/ICECS46596.2019.
stimulation of the muscles. And our study aims to investigate 8965153.
[11] K. Chen, Y. Li, N. Zheng, and Z. Shi, “Research and application
the real-time auxiliary effect of functional electrical stimula- of functional electrical stimulation control system based on adaptive
tion on the gait of stroke patients. However, the long-term EMG modulation,” in Proc. Chin. Intell. Syst. Conf., vol. 804, Y. Jia,
rehabilitation effect was not examined in this study. Subse- W. Zhang, Y. Fu, Z. Yu, and S. Zheng, Eds. Singapore: Springer, 2022,
pp. 412–422, doi: 10.1007/978-981-16-6324-6_42.
quent research will involve evaluating the long-term efficacy
[12] C. Jiang, M. Zheng, Y. Li, X. Wang, L. Li, and R. Song, “Iterative
of the rehabilitation system using the Fugl-Meyer test. adjustment of stimulation timing and intensity during FES-assisted
treadmill walking for patients after stroke,” IEEE Trans. Neural
VI. C ONCLUSION Syst. Rehabil. Eng., vol. 28, no. 6, pp. 1292–1298, Jun. 2020, doi:
This study proposes a multi-channel FES gait rehabilitation 10.1109/TNSRE.2020.2986295.
[13] E. N. Kamavuako, D. Farina, K. Yoshida, and W. Jensen, “Estimation
assistance system based on contralateral surface myoelectric of grasping force from features of intramuscular EMG signals with mir-
control that can effectively assist stroke patients in walking and rored bilateral training,” Ann. Biomed. Eng., vol. 40, no. 3, pp. 648–656,
standardize their gait. A BILSTM model using fewer inputs Mar. 2012, doi: 10.1007/s10439-011-0438-7.
[14] J. Wang, Y. Dai, T. Kang, and X. Si, “Research on gait recognition based
is introduced to control more output channels. A four-channel on lower limb EMG signal,” in Proc. IEEE Int. Conf. Mechatronics
electrical stimulation output board is designed, and a wireless Automat. (ICMA), Takamatsu, Japan, Aug. 2021, pp. 212–217, doi:
myoelectric sensing device is used to collect vastus lateralis 10.1109/ICMA52036.2021.9512759.
[15] P. Wei, J. Zhang, P. Wei, B. Wang, and J. Hong, “Different sEMG
muscle myoelectricity on the non-affected side and modulate and EEG features analysis for gait phase recognition,” in Proc. 42nd
the frequency of the four-channel electrical stimulation on Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Montreal, QC,
the affected side in real time. The reliability of the proposed Canada, Jul. 2020, pp. 1002–1006, doi: 10.1109/EMBC44109.2020.
9175655.
algorithm and the effectiveness of the overall system are
[16] A. Correia, J. M. Martins, and C. P. Santos, “Functional elec-
verified by experiments. The experimental results prove that trical stimulation for gait rehabilitation,” in Proc. 15th Medit.
the proposed prediction model can accurately predict the Conf. Med. Biol. Eng. Comput., vol. 2019, vol. 76, J. Henriques,
complex surface EMG in real time. The constructed system N. Neves, and P. de Carvalho, Eds. Cham, Switzerland: Springer, 2020,
pp. 1954–1966, doi: 10.1007/978-3-030-31635-8_235.
can be effectively used for walking support of hemiplegic [17] R. Luo, S. Sun, X. Zhang, Z. Tang, and W. Wang, “A low-cost end-
patients, and the results presented in this study can provide to-end sEMG-based gait sub-phase recognition system,” IEEE Trans.
a useful theoretical basis and methodological support for FES Neural Syst. Rehabil. Eng., vol. 28, no. 1, pp. 267–276, Jan. 2020, doi:
10.1109/TNSRE.2019.2950096.
rehabilitation products based on EMG control. [18] S. Siami-Namini, N. Tavakoli, and A. S. Namin, “The performance
of LSTM and BiLSTM in forecasting time series,” in Proc. IEEE
R EFERENCES Int. Conf. Big Data (Big Data), Los Angeles, CA, USA, Dec. 2019,
[1] D. Kuriakose and Z. Xiao, “Pathophysiology and treatment of stroke: pp. 3285–3292, doi: 10.1109/BIGDATA47090.2019.9005997.
Present status and future perspectives,” Int. J. Mol. Sci., vol. 21, no. 20, [19] Y. Guo, R. Gravina, X. Gu, G. Fortino, and G.-Z. Yang, “EMG-
p. 7609, Oct. 2020, doi: 10.3390/ijms21207609. based abnormal gait detection and recognition,” in Proc. IEEE Int.
[2] L. M. Weber and J. Stein, “The use of robots in stroke rehabilitation: Conf. Hum.-Mach. Syst. (ICHMS), Rome, Italy, Sep. 2020, pp. 1–6, doi:
A narrative review,” NeuroRehabilitation, vol. 43, no. 1, pp. 99–110, 10.1109/ICHMS49158.2020.9209449.
Jul. 2018, doi: 10.3233/NRE-172408. [20] N. H. Harun, H. M. Desa, and H. A. A. Razak, “Preliminary study of
[3] A. Langeard, L. Bigot, N. Chastan, and A. Gauthier, “Does neuromus- rectus femoris and medial gastrocnemius signals for leg exoskeleton,”
cular electrical stimulation training of the lower limb have functional in Proc. AIP Conf., Strbske Pleso, Slovak Republic, 2019, pp. 1–7, doi:
effects on the elderly? A systematic review,” Experim. Gerontol., vol. 91, 10.1063/1.5118112.
pp. 88–98, May 2017, doi: 10.1016/j.exger.2017.02.070. [21] I. Choi, N. Y. Kim, and C. S. Nam, “A sensorimotor rhythm-based
[4] N. Kapadia, B. Moineau, and M. R. Popovic, “Functional electrical brain–computer interface controlled functional electrical stimulation
stimulation therapy for retraining reaching and grasping after spinal cord for handgrasp rehabilitation,” in Neuroergonomics, C. S. Nam, Ed.
injury and stroke,” Frontiers Neurosci., vol. 14, pp. 1–9, Jul. 2020, doi: Cham, Switzerland: Springer, 2020, pp. 329–349, doi: 10.1007/978-3-
10.3389/fnins.2020.00718. 030-34784-0_17.
LU et al.: MULTI-CHANNEL FES GAIT REHABILITATION ASSISTANCE SYSTEM 3663

[22] M. Moreira and A. P. L. Bó, “Muscle fatigue and the importance of [32] R. H. Gabel and R. A. Brand, “The effects of signal conditioning
electrical stimulation parameters on functional electrical stimulation,” on the statistical analyses of gait EMG,” Electroencephalogr. Clin.
in Proc. Brazilian Congr. Biomed. Eng., vols. 1–70, R. Costa-Felix, Neurophysiol. Evoked Potentials Section, vol. 93, no. 3, pp. 188–201,
J. C. Machado, and A. V. Alvarenga, Eds. Singapore: Springer, 2019, 1994, doi: 10.1016/0168-5597(94)90040-X.
pp. 307–313, doi: 10.1007/978-981-13-2119-1_47. [33] Y. Lu, H. Wang, B. Zhou, C. Wei, and S. Xu, “Continuous and
[23] L.-W. Chou and S. A. Binder-Macleod, “The effects of stimulation simultaneous estimation of lower limb multi-joint angles from sEMG
frequency and fatigue on the force–intensity relationship for human signals based on stacked convolutional and LSTM models,” Expert
skeletal muscle,” Clin. Neurophysiol., vol. 118, no. 6, pp. 1387–1396, Syst. Appl., vol. 203, pp. 1–20, Oct. 2022, doi: 10.1016/j.eswa.2022.
Jun. 2007, doi: 10.1016/j.clinph.2007.02.028. 117340.
[24] L.-W. Chou, T. M. Kesar, and S. A. Binder-Macleod, “Using cus- [34] M. Crepaldi et al., “FITFES: A wearable myoelectrically controlled
tomized rate-coding and recruitment strategies to maintain forces during functional electrical stimulator designed using a user-centered approach,”
repetitive activation of human muscles,” Phys. Therapy, vol. 88, no. 3, IEEE Trans. Neural Syst. Rehabil. Eng., vol. 29, pp. 2142–2152, 2021,
pp. 363–375, Mar. 2008, doi: 10.2522/ptj.20070201. doi: 10.1109/TNSRE.2021.3120293.
[25] M. P. Kadaba, H. K. Ramakrishnan, M. E. Wootten, J. Gainey, [35] K. H. Ew, C. L. Wee, D. G. Zhang, H. Zheng, and K. Y. Zhu, “Effects
G. Gorton, and G. V. B. Cochran, “Repeatability of kinematic, kinetic, of functional electrical stimulation relating to leg movement,” in Proc.
and electromyographic data in normal adult gait,” J. Orthopaedic Res., IEEE Eng. Med. Biol. 27th Annu. Conf., Shanghai, China, Mar. 2005,
vol. 7, no. 6, pp. 849–860, Nov. 1989, doi: 10.1002/jor.1100070611. pp. 6203–6206, doi: 10.1109/IEMBS.2005.1615912.
[26] N. R. Prabhav and S. R. Devasahayam, “Effect of muscle length [36] G. M. Rozanski, A. H. Huntley, L. D. Crosby, A. Schinkel-Ivy,
changes on classification of EMG for prosthesis control,” Int. A. Mansfield, and K. K. Patterson, “Lower limb muscle activ-
J. Biomed. Eng. Technol., vol. 13, no. 2, pp. 117–132, 2013, doi: ity underlying temporal gait asymmetry post-stroke,” Clin. Neu-
10.1504/IJBET.2013.057929. rophysiol., vol. 131, no. 8, pp. 1848–1858, Aug. 2020, doi:
[27] J. R. Potvin and L. R. Bent, “A validation of techniques using surface 10.1016/j.clinph.2020.04.171.
EMG signals from dynamic contractions to quantify muscle fatigue [37] B. A. Knarr, D. S. Reisman, S. A. Binder-Macleod, and J. S. Higginson,
during repetitive tasks,” J. Electromyogr. Kinesiol., vol. 7, no. 2, “Understanding compensatory strategies for muscle weakness
pp. 131–139, Jun. 1997, doi: 10.1016/S1050-6411(96)00025-9. during gait by simulating activation deficits seen post-stroke,”
[28] Z. Tang, K. Zhang, S. Sun, Z. Gao, L. Zhang, and Z. Yang, Gait Posture, vol. 38, no. 2, pp. 270–275, Jun. 2013, doi:
“An upper-limb power-assist exoskeleton using proportional myoelectric 10.1016/j.gaitpost.2012.11.027.
control,” Sensors, vol. 14, no. 4, pp. 6677–6694, Apr. 2014, doi: [38] N. Kumar, N. Kunju, A. Dhawan, A. Kumar, and B. S. Sohi, “Analysis
10.3390/s140406677. of SEMG for treadmill and ground walk of healthy individuals,” in Proc.
[29] S. Guo, Y. Ding, and J. Guo, “Control of a lower limb exoskeleton robot 2nd Int. Conf. Emerg. Trends Eng. Technol., Nagpur, India, Dec. 2009,
by upper limb sEMG signal,” in Proc. IEEE Int. Conf. Mechatronics pp. 46–49, doi: 10.1109/ICETET.2009.32.
Automat. (ICMA), Takamatsu, Japan, Aug. 2021, pp. 1113–1118, doi: [39] A. C. Yep Khoo, Y. Ting Yap, D. Gouwanda, and A. A. Gopalai,
10.1109/ICMA52036.2021.9512775. “Examination of interlimb coordination of human asymmetrical gait,”
[30] C. Frigo, M. Ferrarin, W. Frasson, E. Pavan, and R. Thorsen, “EMG sig- in Proc. IEEE-EMBS Conf. Biomed. Eng. Sci. (IECBES), Sarawak,
nals detection and processing for on-line control of functional electrical Malaysia, Dec. 2018, pp. 680–685, doi: 10.1109/IECBES.2018.
stimulation,” J. Electromyogr. Kinesiol., vol. 10, no. 5, pp. 351–360, 8626691.
2000. [40] C. Morbidoni, A. Cucchiarelli, V. Agostini, M. Knaflitz, S. Fioretti,
[31] T. Bao, Y. Zhao, S. A. R. Zaidi, S. Xie, P. Yang, and Z. Zhang, and F. Di Nardo, “Machine-learning-based prediction of gait events
“A deep Kalman filter network for hand kinematics estimation using from EMG in cerebral palsy children,” IEEE Trans. Neural Syst.
sEMG,” Pattern Recognit. Lett., vol. 143, pp. 88–94, Mar. 2021, doi: Rehabil. Eng., vol. 29, pp. 819–830, 2021, doi: 10.1109/TNSRE.2021.
10.1016/j.patrec.2021.01.001. 3076366.