Evaluating the Viability of Neural Networks for Analysing

Electromyography Data in Home Rehabilitation: Estimating Foot

Progression Angle

Finn Siegel

, Christian Buj

, Ricarda Merfort

, Andreas Hein

and Frerk Müller-Von Aschwege

OFFIS e.V.- Institute for Information Technology, Escherweg 2, Oldenburg, Germany

Universitätsklinikum Aachen, Aachen, Germany

Keywords: Electromyography, Neural Network, Deep Learning, Rehabilitation, Intramedullary Nailing, Femur Shaft

Fracture, Foot Progression Angle.

Abstract: Intramedullary (IM) nailing is a widely accepted treatment for femoral shaft fractures due to its good healing

rate and rapid return to full weight bearing. However, a significant number of patients experience impairments

years after treatment. One possible cause is a malrotation of the femur, resulting in altered foot progression

angles (FPAs), which can lead to changes in gait or persistent pain. To gain a better understanding of

compensation mechanisms and improve rehabilitation strategies, a continuous surface electromyography

(EMG) measurement system worn on vastus lateralis (VL) and vastus medialis (VM) is proposed. To test the

feasibility of this approach, a study is conducted with healthy participants (N=10) simulating different FPA.

The EMG signal was recorded and analysed using a convolutional neural network (CNN). The feasibility

study showed promising results, as the CNN could on average achieve a validation accuracy of 74% in

classifying FPAs as normal, inward (-15°), or outward (+15°). These results show the potential of using EMG

measurements from VL and VM to monitor changes in FPA during rehabilitation. This approach offers the

opportunity to increase our understanding of compensatory mechanisms and improve rehabilitation outcomes

following malrotation caused by IM nailing.

1 INTRODUCTION

The established gold standard for the treatment of

femoral shaft fractures is the use of an intramedullary

(IM) nail. The widespread adoption of this method is

attributed to its compelling properties, including a

high likelihood of fracture union (99%) (Mavrogenis

et al., 2016), a low risk of infection (El Moumni et al.,

2012) and a rapid weight bearing (Rommens &

Hessmann, 2015). However, potential risks include

implant failure (Mavrogenis et al., 2016) or non-

union of the fractured femur (El Moumni et al., 2012).

Despite the relatively low risk associated with IM

nailing, approximately 20% of patients subjected to

this procedure suffer from long-term residual

impairments. These complications can include pain,

hip ossification, altered gait patterns or restricted

https://orcid.org/0000-0002-9319-4304

https://orcid.org/0000-0002-5357-5516

mobility in hip and knee (El Moumni et al., 2012;

Hamahashi et al., 2019). The cause of these

complications remains a topic of ongoing debate.

Surgical factors, including the risk of injuring

surrounding muscle tissue, nerve supply or articular

cartilage may contribute (El Moumni et al., 2012).

Another potential factor could be, that there is no

direct visibility of the femur during surgery making it

difficult to precisely restore rotation and length of the

fractured femur and thus increasing the risk of

malrotation or malpositioning (Jaarsma & van

Kampen, 2004). Such misalignments are defined as

deviations greater than 5° in the frontal or sagittal

plane, 15° in the axial plane, and 2 cm in length (Ricci

et al., 2008). The incidence of such deviations varies

between 22.7% to 28% across studies (Jaarsma & van

Kampen, 2004; Papachristos, 2019; Rommens &

Hessmann, 2015).

132

Siegel, F., Buj, C., Merfort, R., Hein, A. and Aschwege, F.

Evaluating the Viability of Neural Networks for Analysing Electromyography Data in Home Rehabilitation: Estimating Foot Progression Angle.

DOI: 10.5220/0012385100003657

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

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 132-141

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

Irrespective of an identified cause, long-term

residual impairments pose a substantial burden to the

affected patient. Effective rehabilitation, essential for

moderating or even resolving these consequences,

depends on accurate identification of limitations.

After hospitalization, as patients transition to a home-

based care, monitoring is mostly based on subjective

self-assessments, which tend to be inaccurate and can

reduce the quality of rehabilitation (Toogood et al.,

2016).

Previous research by Siegel et al. (2023) suggests

the use of wearable home devices as a strategy to

improve the accuracy of rehabilitation monitoring,

allowing the identification of long-term residual

impairments, thereby providing a basis for the

treating specialist to take adapted countermeasures. In

addition, a continuous monitoring system could

detect possible malrotation and monitor any changes

during rehabilitation (Siegel et al., 2023). Research

by Jaarsma et al. (2004) showed that, on average,

patients are capable of compensating for

approximately 71% of a given malrotation. However,

an enhanced understanding of malrotation

mechanisms could provide further insight into the

compensation strategies and enable clinicians to help

patients cope by training targeted supporting muscles.

This is relevant, as some studies show a high

likelihood for malrotation to be a major source of pain

(Dagneaux et al., 2018).

To gain insight into malrotation, compensation

mechanisms and coping strategies, it is necessary to

continuously measure foot progression angle (FPA).

In a previous paper, the concept of employing a

wearable device, positioned above the knee, equipped

with electromyography (EMG) sensors to measure

the FPA, was introduced (Siegel et al., 2023).

An EMG measurement records biopotentials

when an electrochemical stimulus triggers muscle

fibre (Al-Ayyad et al., 2023). The possibility of using

EMG measurements to draw conclusions about FPA

is based on the premise that alterations in movement

are accompanied by a corresponding change in the

measurable EMG signal (Akuzawa et al., 2017). In

addition, Benedetti el al., 2003 found that altered

muscle contractions, quantifiable through EMG

measurements, may account for a compensation

mechanism (Benedetti et al., 2003).

In order to detect these alterations in muscle

activity, a surface EMG measurement should ideally

record the electrical activity of uniformly active

motor units within one muscle. However, the

resulting EMG signal is subject to many influences,

including fatigue, quantity of active motor units,

firing rates, firing amplitudes, superposition from

surrounding muscles, low-pass characteristics of

surrounding tissue, sensor properties and extraneous

signals such as ambient noise. This complexity makes

it difficult to reliably classify EMG recordings using

basic filter algorithms or feature extraction

methodologies. As a possible answer to this

challenge, deep learning has proven to be a successful

tool (Faust et al., 2018). Especially the usage of

convolutional neural networks (CNNs) has been

proven reliable (Olsson et al., 2019; Yang et al., 2019;

Zia Ur Rehman et al., 2018). CNNs are particularly

suitable for detecting patterns in one-dimensional or

multi-dimensional data due to a high degree of

invariance to translation, scaling, skewing or

distortion. This is possible because each neuron

receives its input from a local receptive field from the

previous layer. Thus, the position of features becomes

less important as long as they maintain their relative

position to each other (Al-Jabery,Khalid et al., 2020),

enabling a classification of variant time series (Zhao

et al., 2017). I.e. Bakircioğlu and Öskurt (2020) used

a CNN to classify EMG recordings of movements

made while gripping six different objects and

achieved 95.9% accuracy. Olsson et al. (2019) used a

CNN, classifying 16 independent states of the hand

recorded using an EMG system and achieved 78.7%

accuracy.

1.1 Aim of this Study

The aim of this study is to evaluate the potential utility

of EMG sensors in improving the reliability of

monitoring FPAs during home-based rehabilitation, a

crucial aspect considering the FPA contributes

significantly in long-term outcome following femoral

shaft fracture treatment.

As of now, EMG measurements have been

successfully used in numerous rehabilitation

applications, such as:

• In neuromuscular rehabilitation, EMG

measurements can be used to quantitatively

assess spasticity as well as monitor treatment

progress (Campanini et al., 2020).

• In post-stroke rehabilitation, EMG

measurements can be used to monitor the

healing process (Simpson et al., 2011) or to

control an exoskeleton aimed to reactivate

paralysed limbs (Nam et al., 2022).

• In orthopaedic rehabilitation, EMG

measurements can be used to evaluate muscle

function, to detect abnormalities or to manage

pain-inducing syndromes during sessions with

specialists (Barton et al., 2013; Benedetti et al.,

2003).

Evaluating the Viability of Neural Networks for Analysing Electromyography Data in Home Rehabilitation: Estimating Foot Progression

Angle

133

Figure 1: Presentation of simulation results, generated with Anybody software. Simulated muscle activity during normal gait

and gait with an intentional alteration (approximately 15° inside or outside) of foot progression angle is displayed. The

overview is shown for vastus medialis and vastus lateralis.

However, until now, EMG measurements have

not been used to monitor a patient’s activity or

malrotation in a home environment.

There are already sensor systems measuring

FPAs, such as inertial measurement units (IMUs) or

pressure sensors, but these have limitations such as

inaccurate results indoors or a dependency on the

footwear (Siegel et al., 2023). An EMG system, worn

on the leg, could potentially overcome these

limitations while increasing data availability, and is

therefore being tested in this study.

Since this study is intended to provide a first

overview of the usability of EMG measurements for

FPA monitoring, it was decided to conduct the tests

on a healthy cohort rather than on patients. The study

will evaluate the following hypotheses:

Hypothesis A: A CNN can classify FPAs of

unknown steps for a single proband, after training on

EMG data obtained from the same proband.

If a CNN is capable do discriminate EMG signals

from different FPAs within a single proband, this

knowledge holds potential to monitor changes in

FPAs during rehabilitation. However, this is limited,

since the proband would need to simulate different

FPAs in order to train such neural network. In

practical clinical scenarios, this data aggregation may

not be feasible, due to the recently treated femur shaft

fracture. Therefore, it is important to investigate,

whether a neural network can be trained using data

from diverse patients and enabling it to classify EMG

signal for different FPAs without prior subject

specific training. To test this the following

hypotheses is formulated:

Hypothesis B: A CNN can classify FPAs for an

unknown proband, after training on EMG data

obtained only different probands.

To test hypotheses A and B, EMG signals of

several probands simulating different FPAs are

recorded and analysed using a CNN.

2 MATERIAL AND METHODS

For the acquisition of EMG data across different

FPAs, ten volunteers were recruited. Exclusion

criteria were adhesive tape and silver allergy,

implanted electrical devices and known deformities

of the lower limb. The gender distribution was 50%

male to 50% female with a mean age of 36.5 years

(±14.3 years).

2.1 Sensor Placement

To ensure optimal sensor placement, aligned to answer

the hypotheses, literature was reviewed to find poten-

tial correlations between lower limb muscle activity

and FPA. A simulation was used to verify the results.

Mohammad and Elsais (2020) investigated the

correlation between EMG signal amplitude and hip

rotation in male runners. They found such a

relationship for Vastus Lateralis (VL), Vastus Medius

(VM) and Gluteus Maximus. As the overall goal,

outlined in earlier work (Siegel et al., 2023), is to

wear a sensor array positioned above the knee, the

electrical activity of VL and VM are promising

muscles for conclusions to be drawn about FPAs.

HEALTHINF 2024 - 17th International Conference on Health Informatics

134

Figure 2: Exemplary image from the study displaying

placement of EMG sensors on VL and VM. The GaitRite

mat, used for FPA verification, is also visible.

To further confirm that change in FPA induces

change in muscle activity for VL and VM, a

simulation was performed in collaboration with

university hospital in Aachen, Germany. A volunteer

was fitted with a MTw Awinda motion tracker system

(Paulich et al., 2018) and walks with different FPAs

were conducted. The FPA was varied between 15°

inward, normal and 15° outward. To derive muscle

activity from the collected data, the AnyBody

Modelling System (Paul & Doweidar, 2023) was used

in combination with the AnyBody Managed Model

Repository. This approach allows an inverse

dynamics analysis to be performed, based on a third-

order polynomial muscle recruitment criterion, which

produces a simulation of the electrical activity in the

lower limb muscles during walks. The results are

shown in figure 1. This figure displays muscle

activity during gait with normal FPA compared to

gait with an inward or outward FPA. The simulated

activity is shown for VM and VL. It is immediately

noticeable, that the shape of the curves for normal and

modified FPA are distinctly different. To quantify

these observations, the integral of the curves was

calculated (python library: numpy.trapz (Harris et al.,

2020)), displayed in table 1. The results show

variation in the area under the curve for normal FPAs

compared to modified FPAs. The differences are

particularly significant for outside FPAs in

comparison to inside FPAs. The simulation supports

the choice of using VL and VM as EMG

measurement points to detect differences in FPA.

In conclusion it was decided to place the EMG

sensors on VL and VM. The European

recommendations for sensors and sensor placement

for EMG (Hermens et al., n.d.) was used as a guide to

ensure optimal placement of the sensors on VL and

VM, minimising superposing of signals by

surrounding muscles. To further improve signal

Table 1: To quantify figure 1, the area under the plotted

curves is calculated using the trapezoid method (python

library: numpy.trapz) and the results are shown in this table.

quality, the skin was shaved and cleaned prior to

sensor placement. An example of placed sensors is

given in figure 2.

2.2 Signal Acquisition

The EMG signal was recorded using the Delsys

Trigno-Wireless-Biofeedback System (Delsys, n.d.).

This system consists of a base station that wirelessly

collects data from individual sensors. Each sensor is

capable of collecting data at a frequency of 4 kHz

with a bandwidth of 20-450 Hz and an input range of

11 mV.

For the purpose of supervised learning, it is

necessary to label EMG data recordings.

Consequently, a GaitRite mat was used to record the

FPAs. The mat is manufactured by CIR Systems

represents a gold standard in gait analysis. 36 864

pressure sensors evenly distributed of over a length of

914 cm allows steps to be recorded and a gait profile

to be created. This profile includes the FPA for each

step executed. A section of the mat can be seen in the

figure 2.

In order to conduct this study, each proband had

to perform a total of 45 walks along the entire length

of the GaitRite. 15 normal walks, 15 walks with

outward FPAs and 15 walks with inward FPAs. For

each simulated malrotation, the participants were

asked to change their FPA by -15° inward or +15°

outward. Prior to the study, foot positions were

trained using the GaitRite mat. During the study any

steps deviating by more than ±8° from proposed FPA

were removed from the dataset. Only one foot was

varied during the different walks. The side to be

varied was freely chosen by the proband, the choice

remained consistent throughout the study. A

metronome was used to ensure uniform walking

speed during different walks. In order to merge EMG

data with GaitRite data, software was developed to

record both systems simultaneously. Both data

streams were synchronised by an output signal

generated by the GaitRite system.

Evaluating the Viability of Neural Networks for Analysing Electromyography Data in Home Rehabilitation: Estimating Foot Progression

Angle

135

2.2.1 Data Preprocessing

Data preprocessing is performed according to a

general data preparation paradigm (Al-Jabery et al.,

2020). During the study, walks across the GaitRite

mat were recorded alongside the corresponding EMG

signals, resulting in 15 datasets per class (inward,

outward and normal FPA). However, this quantity

proved insufficient to train a supervised deep learning

algorithm (Alwosheel et al., 2018). To increase the

size of each class, the walks are divided into

individual steps. For this purpose, software was

developed that extracts individual steps based on

EMG peak detection and assigns them to the

appropriate FPA class. This results in a dataset for

each proband containing the EMG signal for VL and

VM and the corresponding FPA for each step.

To extract non-stationary properties from the

EMG signal, time windowing is performed (Zha et

al., 2021). Initially, the EMG signal of one step spans

over a duration of one second. This can be contracted,

since VL and VM are only active for approximately

20% to 25% of the time during one gait cycle (Róisín

Howard, 2017). The average duration of a gait cycle

is around one second (Murray et al., 1964), enabling

the EMG signal be to contracted to a duration of 250

ms. As data was collected at 4 kHz, the EMG signal

is truncated to a time window of 1000 data points

(250ms). The next step is a high and low pass filtering

(Morbidoni et al., 2019), already conducted by the

Table 2: This table displays the distribution of steps

generated in this study across different subjects and FPAs,

showing the class sizes used to train the neural network.

sensor. To filter motion artifacts, a low-pass filter

with a cutoff frequency of 20 Hz is used. A high-pass

filter with a cutoff frequency of 450 Hz is applied, as

not much additional information is available above

this frequency (Bakircioğlu & Özkurt, 2020). This is

followed by a rectification of the data, enhancing the

chances of successful training of deep learning

algorithms (Li et al., 2011). Next, a Fast Fourier

Transformation (FFT) is performed creating

additional input features and enhancing information

density (Yang et al., 2019). Finally, data is

normalised using the peak-dynamic method,

requiring each data point to be divided by the

maximum value. While this method results in a loss

of information regarding the degree of muscle

activation, it enhances the comparability between

probands.

In Conclusion, a matrix is generated containing

both a time series and a frequency series, for each

labelled step and each sensor. Combining the

measurements for VL and VM results in a matrix of

Table 3: This table shows the structure of the CNN used. The optimizer adam and sparse categorical crossentropy were used

for training.

HEALTHINF 2024 - 17th International Conference on Health Informatics

136

Table 4: Representation of the accuracy achieved for individual subjects using a CNN for classifying the classes normal,

inside and outside FPAs. The network was trained and evaluated three times. The average accuracy and the corresponding

standard deviation are also presented.

four features with 1000 data points times the number

of steps. In this study, a total of 186-238 steps were

recorded, per participant. Resulting in 2146 steps

available to train the CNN, see table 2. This is a

relatively modest dataset size for the application of

deep learning (Alwosheel et al., 2018), but the

purpose of this study is to provide an initial insight in

the possibility of determining FPAs using EMG

measurements in conjunction with deep learning

evaluations and therefore declared acceptable for this

feasibility study.

2.2.2 Used Network

The structure of the CNN used is shown in the table

3. The network is built using TensorFlow (Martín

Abadi et al., 2015) and Keras (Chollet & others,

2015) libraries in Python. To obtain reliable results,

each run was performed three times and the average

validation accuracy is taken as the result.

3 RESULTS

In the following the results gained from the analysis

of the study data are presented in relation to the tested

hypothesis.

3.1 Results for Testing Hypothesis A

H: A CNN can classify FPAs of unknown steps for a

single proband, after training on EMG data obtained

from the same proband.

To test this hypothesis, data from the gait study

obtained by each proband individually was used to

train the CNN. The labelled data was combined,

randomly mixed and 85% was used for training and

the remaining 15% served as validation data. To

Figure 3: Confusion matrix, showing the result for the best

prediction. The predicted value is shown against the true

value.

Figure 4: Confusion matrix, showing the result for the least

successful prediction. The predicted value is shown against

the true value.

Evaluating the Viability of Neural Networks for Analysing Electromyography Data in Home Rehabilitation: Estimating Foot Progression

Angle

137

Table 5: Presentation of the interproband validation accuracy achieved using a CNN to classify the classes normal, inside and

outside FPAs. Shown is the average validation accuracy.

ensure a reliable conclusion, each training iteration

was performed three times. The results are presented

in table 4.

Steps are classified into three classes (normal,

internal and external rotated FPAs) with an average

classification accuracy of 74.2% (±10.4%). This

performance exceeds chance level of 33. 3



% .

Additionally, a confusion matrix is displayed, for the

most successful and the least successful

classification, see figure 3.

The result suggests that a CNN can learn and

discriminate features within the EMG signal allowing

conclusions to be drawn about FPAs. However, the

high variance of 10.4% indicates inconsistency in the

quality of features identified by the CNN between

probands.

3.2 Results for Testing Hypothesis B

H: A CNN can classify FPAs for an unknown

proband, after training on EMG data obtained from

different probands.

To test this hypothesis, datasets from all probands

excluding one for validation were combined and used

to train a CNN. This process was repeated, ensuring

each proband’s data was tested against the combined

majority. The results are shown in table 5. It can be

seen that a CNN, trained on a whole population, can

distinguish validation steps an average accuracy of

40% across inward, outward and normal FPAs. The

results indicate a limited reliability for a classification

of unknown EMG data recorded from different FPAs.

4 DISCUSSION

In this study, each class (normal, inside and outside

FPA) contains 1326-1488 trails (for VL and VM

combined), which, in the context of deep learning,

accounts for a relatively small dataset (Alwosheel et

al., 2018). However, when working with EMG

measurements, the availability of data is limited by

the number of times a person can repeat a specific

movement. This limitation restricts the size of

available datasets, which needs to be considered when

working with deep neural networks. Nevertheless,

researchers have shown that small datasets can be used

successfully, i.e. Grag et al. (2021) used three classes

of EMG recordings and a total of 1575 trails while

achieving an accuracy of 85.44%.

The inclusion of 10 probands, as in this study, is

in line with the approach of other researchers, when

experimenting with EMG data. I.e. Rehman et al.

(2018) collected data from seven healthy probands

and Bakircioğlu and Öskurt (2020) had five probands

enrolled in their study.

4.1 Discussion of Hypothesis A

H: A CNN can classify FPAs of unknown steps for a

single proband, after training on EMG data obtained

from the same proband.

This study has shown a CNN can learn features from

EMG recordings of VL and VM to distinguish

between outward, inward and normal FPAs with an

average success rate of 74.2%. The standard deviation

of 10.4% reflects the high variance of the EMG

signal, which has also been reported by other

researchers (Rane et al., 2019). The variability of the

EMG signal can be attributed to its inherent nature,

which is non-stationary, non-linear, stochastic, and

unpredictable (Geng et al., 2016). At the same time,

the characteristics of the sensor play a role, as the

signal varies depending on the position relative to the

muscle and the quality of the contact with the skin. In

addition, the signal is prone to noise, including

instrument noise, ambient noise, motion artefacts, and

signal instability (Reaz et al., 2006).

The result of this part of the study is in line with

results of other studies, i.e. Tryon et al. (2021)

achieved an accuracy of 74.7% in discriminating

EMG signals into three classes related to elbow

flexion while holding different weights.

It is important to note, when dealing with hand

gestures using a CNN, results tend to be significantly

better. For instance, Lee at al. (2020) achieved an

accuracy of 94% when discriminating between ten

gesture classes. The differences in performance may

be due to the availability of distinct movements,

whereas this study focuses on detecting small changes

HEALTHINF 2024 - 17th International Conference on Health Informatics

138

in movement sequences, which are easily masked by

the noise of the EMG signal.

An improvement in results could be achieved by

using CNNs in combination with other deep learning

algorithms. For example, by connecting CNNs to

bidirectional LSTM networks. Karnam et al. (2022)

were able to improve the accuracy of classifying

EMG recordings of hand gestures by up to 18.7%

compared to state-of-the-art models.

Another way to improve the accuracy of the CNN

is using transfer learning. This involves pre-training

the network on subjects with comparable data

recorded from other subjects followed by training on

target data. Soroushmojdehi et al. (2022) showed that

this methodology can improve the accuracy of a

CNN, when predicting hand movements based on

EMG data, up to 10%.

4.2 Discussion of Hypothesis B

H: A CNN can classify FPAs for an unknown

proband, after training on EMG data obtained from

different probands.

The result of this hypothesis testing shows an average

accuracy of 40.3%, barely surpassing chance level

(33.33%). A major contributing factor is the high

interpatient variability. This high variability has

already been reported by Anders et al. (2019), who

demonstrated substantial interindividual variability

and Guidetti et al. (1996) found significant variation

between subjects.

Furthermore, the interpatient comparison results

are consistent with findings in existing literature. In

this study, three classes of FPAs were classified with

up to 46.8% validation accuracy, see table 4. This

performance is comparable to that of Castellini’s team,

who achieved an accuracy of 51.7% for three classes in

an interpatient evaluation (Castellini et al., 2009).

One way to improve the results could be to use the

normal gait pattern of a subject under investigation as

calibration followed by detecting changes in FPAs

with the help of a trained CNN. Cano et al. (2022)

showed, that the accuracy of predicting high blood

pressure in unknown subjects could be increased by

up to 30% this way.

5 CONCLUSIONS

The aim of this study was to provide initial insights

into the potential utility of EMG sensors in improving

the reliability of FPA monitoring during home

rehabilitation. It has been demonstrated that EMG

measurements, evaluated by a CNN trained on an

individual proband, can be used to classify between

inward, outward and normal FPAs with an average

validation accuracy of 70.4%. In conclusion, while

the results show that such a system is not yet ready

for use as a medical device, they highlight the

potential and need for further research into this

approach.

The major goal for the future is to develop a user-

friendly measuring device capable of precisely

detecting changes in FPA, providing essential data for

the recovery process. The next steps on this path

include minimising the variance between different

patients. The use of an EMG sensor array is one

possible solution for this, as it allows the

determination of the sensor with the optimal signal

quality, thus reducing the need for precise sensor

placement. In addition, increasing the size of the data

set is a critical factor. The possibility to integrate

more steps could significantly increase the accuracy

of a neural network. other optimisation approaches

include combining different deep learning algorithms

and testing the usability of transfer learning.

To the best of our knowledge, this study

represents the first instance of utilizing EMG

measurements in combination with CNNs to provide

insight into FPA.

ACKNOWLEDGEMENTS

This work is founded by the German Federal Ministry

of Education and Research (BMBF) (FKZ:

01IS21085) and is part of the ITEA Secure-e-Health

project.

REFERENCES

Akuzawa, H., Imai, A., Iizuka, S., Matsunaga, N., &

Kaneoka, K. (2017). The influence of foot position on

lower leg muscle activity during a heel raise exercise

measured with fine-wire and surface EMG. Physical

Therapy in Sport, 28, 23–28. https://doi.org/10.1016/

j.ptsp.2017.08.077

Al-Ayyad, M., Owida, H. A., De Fazio, R., Al-Naami, B., &

Visconti, P. (2023). Electromyography Monitoring

Systems in Rehabilitation: A Review of Clinical

Applications, Wearable Devices and Signal Acquisition

Methodologies. Electronics, 12(7), 1520. https://doi.org/

10.3390/electronics12071520

Al-Jabery,Khalid, Obafemi-Ajayi, Tayo, Olbricht, Gayla, &

Wunsch, Donald. (2020). Computational learning

approaches to data analytics in biomedical applications.

Academic press.

Evaluating the Viability of Neural Networks for Analysing Electromyography Data in Home Rehabilitation: Estimating Foot Progression

Angle

139

Alwosheel, A., Van Cranenburgh, S., & Chorus, C. G.

(2018). Is your dataset big enough? Sample size

requirements when using artificial neural networks for

discrete choice analysis. Journal of Choice Modelling,

28, 167–182. https://doi.org/10.1016/j.jocm.2018.07.002

Anders, J. P. V., Smith, C. M., Keller, J. L., Hill, E. C.,

Housh, T. J., Schmidt, R. J., & Johnson, G. O. (2019).

Inter- and Intra-Individual Differences in EMG and

MMG during Maximal, Bilateral, Dynamic Leg

Extensions. Sports, 7(7), 175. https://doi.org/10.3390/

sports7070175

Bakircioğlu, K., & Özkurt, N. (2020). Classification of Emg

Signals Using Convolution Neural Network.

International Journal of Applied Mathematics

Electronics and Computers, 8(4), 115–119.

https://doi.org/10.18100/ijamec.795227

Barton, C. J., Lack, S., Malliaras, P., & Morrissey, D. (2013).

Gluteal muscle activity and patellofemoral pain

syndrome: A systematic review. British Journal of Sports

Medicine, 47(4), 207–214. https://doi.org/10.1136/

bjsports-2012-090953

Benedetti, M. G., Catani, F., Bilotta, T. W., Marcacci, M.,

Mariani, E., & Giannini, S. (2003). Muscle activation

pattern and gait biomechanics after total knee

replacement. Clinical Biomechanics, 18(9), 871–876.

https://doi.org/10.1016/S0268-0033(03)00146-3

Campanini, I., Disselhorst-Klug, C., Rymer, W. Z., &

Merletti, R. (2020). Surface EMG in Clinical Assessment

and Neurorehabilitation: Barriers Limiting Its Use.

Frontiers in Neurology, 11, 934. https://doi.org/

10.3389/fneur.2020.00934

Cano, J., Fácila, L., Gracia-Baena, J. M., Zangróniz, R.,

Alcaraz, R., & Rieta, J. J. (2022). The Relevance of

Calibration in Machine Learning-Based Hypertension

Risk Assessment Combining Photoplethysmography and

Electrocardiography. Biosensors, 12(5), 289.

https://doi.org/10.3390/bios12050289

Castellini, C., Fiorilla, A. E., & Sandini, G. (2009). Multi-

subject/daily-life activity EMG-based control of

mechanical hands. Journal of NeuroEngineering and

Rehabilitation, 6(1), 41. https://doi.org/10.1186/1743-

0003-6-41

Chollet, F. & others. (2015). Keras. https://keras.io

Dagneaux, L., Allal, R., Pithioux, M., Chabrand, P., Ollivier,

M., & Argenson, J.-N. (2018). Femoral malrotation from

diaphyseal fractures results in changes in patellofemoral

alignment and higher patellofemoral stress from a finite

element model study. The Knee, 25(5), 807–813.

https://doi.org/10.1016/j.knee.2018.06.008

Delsys. (n.d.). Trigno Wireless Biofeedback System.

https://www.delsys.com/downloads/USERSGUIDE/trig

no/wireless-biofeedback-system.pdf

El Moumni, M., Voogd, E. H., Ten Duis, H. J., & Wendt, K.

W. (2012). Long-term functional outcome following

intramedullary nailing of femoral shaft fractures. Injury,

43(7), 1154–1158. https://doi.org/10.1016/j.injury.20

12.03.011

Faust, O., Hagiwara, Y., Hong, T. J., Lih, O. S., & Acharya,

U. R. (2018). Deep learning for healthcare applications

based on physiological signals: A review. Computer

Methods and Programs in Biomedicine, 161, 1–13.

https://doi.org/10.1016/j.cmpb.2018.04.005

Garg, N., Balafrej, I., Beilliard, Y., Drouin, D., Alibart, F., &

Rouat, J. (2021). Signals to Spikes for Neuromorphic

Regulated Reservoir Computing and EMG Hand Gesture

Recognition. International Conference on Neuromorphic

Systems 2021, 1–8. https://doi.org/10.1145/3477145.

3477267

Geng, W., Du, Y., Jin, W., Wei, W., Hu, Y., & Li, J. (2016).

Gesture recognition by instantaneous surface EMG

images. Scientific Reports, 6(1), 36571. https://doi.org/

10.1038/srep36571

Guidetti, L., Rivellini, G., & Figura, F. (1996). EMG patterns

during running: Intra- and inter-individual variability.

Journal of Electromyography and Kinesiology, 6(1), 37–

48. https://doi.org/10.1016/1050-6411(95)00015-1

Hamahashi, K., Uchiyama, Y., Kobayashi, Y., Ebihara, G.,

Ukai, T., & Watanabe, M. (2019). Clinical outcomes of

intramedullary nailing of femoral shaft fractures with

third fragments: A retrospective analysis of risk factors

for delayed union. Trauma Surgery & Acute Care Open,

4(1), e000203. https://doi.org/10.1136/tsaco-2018-

000203

Harris, C. R., Millman, K. J., Walt, S. J. van der, Gommers,

R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor, J.,

Berg, S., Smith, N. J., Kern, R., Picus, M., Hoyer, S.,

Kerkwijk, M. H. van, Brett, M., Haldane, A., Río, J. F.

del, Wiebe, M., Peterson, P., … Oliphant, T. E. (2020).

Array programming with NumPy. Nature, 585(7825),

357–362. https://doi.org/10.1038/s41586-020-2649-2

Hermens, Freriks, Merletti, Rau, Disselhorst-Klug, &

Stegeman. (n.d.). SENIAM (Surface ElectroMyoGraphy

for the Non-Invasive Assessment of Muscles) project.

http://www.seniam.org/

Jaarsma, R. L., Ongkiehong, B. F., Grüneberg, C.,

Verdonschot, N., Duysens, J., & van Kampen, A. (2004).

Compensation for rotational malalignment after

intramedullary nailing for femoral shaft fractures. Injury,

35(12), 1270–1278. https://doi.org/10.1016/j.injury.20

04.01.016

Jaarsma, R. L., & van Kampen, A. (2004). Rotational

malalignment after fractures of the femur. The Journal of

Bone and Joint Surgery. British Volume, 86-B(8), 1100–

1104. https://doi.org/10.1302/0301-620X.86B8.15663

Karnam, N. K., Dubey, S. R., Turlapaty, A. C., & Gokaraju,

B. (2022). EMGHandNet: A hybrid CNN and Bi-LSTM

architecture for hand activity classification using surface

EMG signals. Biocybernetics and Biomedical

Engineering, 42(1), 325–340. https://doi.org/10.1016/

j.bbe.2022.02.005

Lee, S., Sung, M., & Choi, Y. (2020). Wearable fabric sensor

for controlling myoelectric hand prosthesis via

classification of foot postures. Smart Materials and

Structures, 29(3), 035004. https://doi.org/10.1088/1361-

665X/ab6690

Li, G., Li, Y., Yu, L., & Geng, Y. (2011). Conditioning and

Sampling Issues of EMG Signals in Motion Recognition

of Multifunctional Myoelectric Prostheses. Annals of

Biomedical Engineering, 39(6), 1779–1787.

https://doi.org/10.1007/s10439-011-0265-x

Martín Abadi, Ashish Agarwal, Paul Barham, Eugene

Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado,

Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay

Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey

HEALTHINF 2024 - 17th International Conference on Health Informatics

140

Irving, Michael Isard, Jia, Y., Rafal Jozefowicz, Lukasz

Kaiser, Manjunath Kudlur, … Xiaoqiang Zheng. (2015).

TensorFlow: Large-Scale Machine Learning on

Heterogeneous Systems. https://www.tensorflow.org/

Mavrogenis, A. F., Panagopoulos, G. N., Megaloikonomos,

P. D., Igoumenou, V. G., Galanopoulos, I., Vottis, C. T.,

Karabinas, P., Koulouvaris, P., Kontogeorgakos, V. A.,

Vlamis, J., & Papagelopoulos, P. J. (2016).

Complications After Hip Nailing for Fractures.

Orthopedics, 39(1). https://doi.org/10.3928/01477447-

20151222-11

Mohammad, W. S., & Elsais, W. M. (2020). Association

Between Hip Rotation and Activation of the Quadriceps

and Gluteus Maximus in Male Runners. Orthopaedic

Journal of Sports Medicine, 8(11), 232596712096280.

https://doi.org/10.1177/2325967120962802

Morbidoni, C., Cucchiarelli, A., Fioretti, S., & Di Nardo, F.

(2019). A Deep Learning Approach to EMG-Based

Classification of Gait Phases during Level Ground

Walking. Electronics, 8(8), 894. https://doi.org/

10.3390/electronics8080894

Murray, M. P., Drought, A. B., & Kory, R. C. (1964).

Walking Patterns Of Normal Men. The Journal of Bone

and Joint Surgery. American Volume, 46, 335–360.

Nam, C., Rong, W., Li, W., Cheung, C., Ngai, W., Cheung,

T., Pang, M., Li, L., Hu, J., Wai, H., & Hu, X. (2022). An

Exoneuromusculoskeleton for Self-Help Upper Limb

Rehabilitation After Stroke. Soft Robotics, 9(1), 14–35.

https://doi.org/10.1089/soro.2020.0090

Olsson, A. E., Sager, P., Andersson, E., Björkman, A.,

Malešević, N., & Antfolk, C. (2019). Extraction of Multi-

Labelled Movement Information from the Raw HD-

sEMG Image with Time-Domain Depth. Scientific

Reports, 9(1), 7244. https://doi.org/10.1038/s41598-019-

43676-8

Papachristos, I. V. (2019). Complications of Femoral

Intramedullary Nailing: What should the Surgeon

Remember? . . EC, 7.

Paul, G., & Doweidar, M. H. (Eds.). (2023). Digital human

modeling and medicine: The digital twin. Academic

Press, an imprint of Elsevier.

Paulich, M., Schepers, M., Rudigkeit, N., & Bellusci, G.

(2018). Xsens MTw Awinda: Miniature Wireless Inertial-

Magnetic Motion Tracker for Highly Accurate 3D

Kinematic Applications. https://doi.org/10.13140/

RG.2.2.23576.49929

Rane, L., Ding, Z., McGregor, A. H., & Bull, A. M. J. (2019).

Deep Learning for Musculoskeletal Force Prediction.

Annals of Biomedical Engineering, 47(3), 778–789.

https://doi.org/10.1007/s10439-018-02190-0

Reaz, M. B. I., Hussain, M. S., & Mohd-Yasin, F. (2006).

Techniques of EMG signal analysis: Detection,

processing, classification and applications. Biological

Procedures Online, 8(1), 11–35. https://doi.org/10.1251/

bpo115

Ricci, W. M., Schwappach, J., Tucker, M., Coupe, K.,

Brandt, A., Sanders, R., & Leighton, R. (2008).

Trochanteric versus Piriformis Entry Portal for the

Treatment of Femoral Shaft Fractures. Journal of

Orthopaedic Trauma, 22, S9–S13. https://doi.org/

10.1097/01.bot.0000248472.53154.14

Róisín Howard. (2017). The application of data analysis

methods for surface electromyography in shot putting

and sprinting. https://doi.org/10.13140/RG.2.2.1590

7.04640

Rommens, P. M., & Hessmann, M. H. (Eds.). (2015).

Intramedullary Nailing. Springer London.

https://doi.org/10.1007/978-1-4471-6612-2

Siegel, F., Buj, C., Schwanbeck, R., Petersik, A., Hoffmann,

U., Kemper, J., Hildebrand, F., Kobbe, P., Eschweiler, J.,

Greven, J., Merfort, R., Freimann, C., Schwaiger, A., &

Aschwege, F. (2023). Concept for General

Improvements in the Treatment of Femoral Shaft

Fractures with an Intramedullary Nail: Proceedings of

the 16th International Joint Conference on Biomedical

Engineering Systems and Technologies, 360–367.

https://doi.org/10.5220/0011679100003414

Simpson, K. M., Munro, B. J., & Steele, J. R. (2011).

Backpack load affects lower limb muscle activity

patterns of female hikers during prolonged load carriage.

Journal of Electromyography and Kinesiology, 21(5),

782–788. https://doi.org/10.1016/j.jelekin.2011.05.012

Soroushmojdehi, R., Javadzadeh, S., Pedrocchi, A., &

Gandolla, M. (2022). Transfer learning in hand

movement intention detection based on surface

electromyography signals. Frontiers in Neuroscience,

16, 977328. https://doi.org/10.3389/fnins.2022.977328

Toogood, P. A., Abdel, M. P., Spear, J. A., Cook, S. M.,

Cook, D. J., & Taunton, M. J. (2016). The monitoring of

activity at home after total hip arthroplasty. The Bone &

Joint Journal, 98-B(11), 1450–1454. https://doi.org/

10.1302/0301-620X.98B11.BJJ-2016-0194.R1

Tryon, J., & Trejos, A. L. (2021). Evaluating Convolutional

Neural Networks as a Method of EEG–EMG Fusion.

Frontiers in Neurorobotics, 15, 692183. https://doi.org/

10.3389/fnbot.2021.692183

Yang, W., Yang, D., Liu, Y., & Liu, H. (2019). EMG Pattern

Recognition Using Convolutional Neural Network with

Different Scale Signal/Spectra Input. International

Journal of Humanoid Robotics, 16(04), 1950013.

https://doi.org/10.1142/S0219843619500130

Zha, X., Wehbe, L., Sclabassi, R. J., Mace, Z., Liang, Y. V.,

Yu, A., Leonardo, J., Cheng, B. C., Hillman, T. A., Chen,

D. A., & Riviere, C. N. (2021). A Deep Learning Model

for Automated Classification of Intraoperative

Continuous EMG. IEEE Transactions on Medical

Robotics and Bionics, 3(1), 44–52.

https://doi.org/10.1109/TMRB.2020.3048255

Zhao, B., Lu, H., Chen, S., Liu, J., & Wu, D. (2017).

Convolutional neural networks for time series

classification. Journal of Systems Engineering and

Electronics, 28(1), 162–169. https://doi.org/10.21629/

JSEE.2017.01.18

Zia Ur Rehman, M., Waris, A., Gilani, S., Jochumsen, M.,

Niazi, I., Jamil, M., Farina, D., & Kamavuako, E. (2018).

Multiday EMG-Based Classification of Hand Motions

with Deep Learning Techniques. Sensors, 18(8), 2497.

https://doi.org/10.3390/s18082497

Evaluating the Viability of Neural Networks for Analysing Electromyography Data in Home Rehabilitation: Estimating Foot Progression

Angle

141