Impact of Feature Extraction Optimization on Machine Learning Models

for sEMG-Based Prosthesis Control

Ricardo Henrique Avelar Matheus and Maria Claudia F. Castro

Department of Electrical Engineering, Centro Universit

ario FEI, Brazil

Keywords:

sEMG, Machine Learning, Feature Extraction, LDA, RMS, Amputees.

Abstract:

One of the most signiﬁcant challenges to the quality of life for amputees is the development of prostheses

that can closely simulate the capabilities of the lost limb. One possible solution to this problem is myo-

electric prostheses, which are devices that use myolectric signals as users’ intention to perform independent

movements. This study aims to investigate how optimizing feature extraction methods can impact the perfor-

mance of machine learning models in recognizing surface electromyogram (sEMG) signals from amputees.

The LibEMG library in Python, which offers a simple and robust API for developing sEMG-based projects,

was used alongside the DB8 dataset from the NINAPRO public database, which promotes machine-learning

research in human, robotic, and prosthetic hands. A total of twelve feature extraction methods and seven dif-

ferent classiﬁers were tested. The results showed the best mean accuracy of 79.18% using a Random Forest

classiﬁer with a set of eleven time and frequency domain features, considering the data of an amputee with

experience in using myoelectric prostheses. However, the most affected models by feature optimization were

KNN, MLP, and SVM, with accuracy improvements up to 69.28%.

1 INTRODUCTION

One of the most discussed social issues today is ac-

cessibility and quality of life, particularly for individ-

uals with disabilities. Speciﬁcally, for people with

upper-limb amputations, one of the most signiﬁcant

challenges is the development of prostheses that can

closely simulate the capabilities of the lost limb. One

possible solution to this problem is myoelectric pros-

theses, devices that use myolectric signals as users in-

tend to perform independent movements, mimicking

the functions of a healthy limb (Andrade, 2007).

The common and affordable myoelectric prosthe-

ses available today are primarily based on the tim-

ing and number of electrical pulses generated by the

user within a speciﬁc period, limiting their ability

to perform independent movements. Machine learn-

ing techniques can enhance the number of move-

ments required to make these prostheses increas-

ingly autonomous, natural, and more intuitive to the

user. Generally, classiﬁcation techniques are em-

ployed to recognize patterns in surface electromyo-

graphy (sEMG) signals from various muscles so that

the prosthesis can identify the user’s intended move-

https://orcid.org/0000-0002-2751-0014

ment and respond according to the signal.

One of the ongoing research efforts in this ﬁeld is

the development of more efﬁcient recognition mod-

els with higher accuracy for as many movements as

possible. However, most of the literature is applied

to health volunteer sEMG data (Sultana et al., 2023;

Song et al., 2023). Atzori et al. (2023) compared

the classiﬁcation of 53 movements between one indi-

vidual with a transradial amputation and a database

of individuals with intact upper limbs. Using a K

— Nearest Neighbors (k-NN) classiﬁer resulted in an

average accuracy of 61.51% for the amputee, com-

pared to 80.16% for the non-amputees. In the same

study, another test reducing the number of movements

to 13 resulted in 100% accuracy, exceeding those re-

ported in the literature by the author, such as 95%

accuracy for six movements using Principal Compo-

nent Analysis (PCA) and Support Vector Machine

(SVM) (Castellini et al., 2009), 84.4% accuracy for

ten movements using time-domain features and Lin-

ear Discriminant Analysis (LDA) (Li et al., 2010),

and 87.8% accuracy for twelve movements also us-

ing time-domain features but with MLP (Tenore et al.,

2009). Using MyoArmband, a device with a low-

frequency sEMG acquisition, Cognolato et al. (2018)

found signiﬁcant accuracy discrepancies reached by

Matheus, R. H. A. and Castro, M. C. F.

Impact of Feature Extraction Optimization on Machine Learning Models for sEMG-Based Prosthesis Control.

DOI: 10.5220/0013173300003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 907-913

ISBN: 978-989-758-731-3; ISSN: 2184-4305

907

transradial amputees, ranging from 50% to 97.2%,

where a user of a myoelectric prosthesis achieved the

best.

The performance discrepancies between health

and amputees emphasize the importance of includ-

ing amputee data in developing these models. In the

literature, we also observed a signiﬁcant variance in

the accuracies achieved with amputee data. Factors

such as user experience and training with myoelec-

tric prostheses, the number of movement classes, type

of movement, classiﬁers used, and feature extraction

methods also inﬂuence the classiﬁcation outcome.

Thus, the general objective of this study was to an-

alyze a classiﬁcation model using various classiﬁers

and feature extraction methods to achieve the best re-

sults in terms of accuracy, precision, recall, and F1

score. For this purpose, the dataset provided by the

NINAPRO online library, which included nine move-

ment classes and data from a single transradial am-

putee, was used.

2 METHODS

The proposed methodology is shown in Figure 1.

Figure 1: Methodology Workﬂow.

2.1 Dataset

The DB8 dataset, available in (NINAPRO multimodal

database, 2024) and described in detail in (Krasoulis

et al., 2019), consists of sEMG, IMU, and kinematic

data from 12 subjects, two of whom have transradial

amputations. For this study, only sEMG data from

both amputee subjects were considered. Both subjects

were male, with a right-hand amputation; subject 11

(S11) was 30 years old with no experience using my-

oelectric prostheses, while subject 12 (S12) was 56

with 2 years of experience using myoelectric prosthe-

ses.

Data acquisition was performed using 16 Delsys

Trigno IM Wireless sEMG system sensors placed on

the user’s forearm to capture the surface EMG at a

sampling frequency of 2kHz. No speciﬁc muscle was

chosen as the focus of acquisition. The positions of

the electrodes are shown in Figure 2.

Figure 2: sEMG sensors location (Krasoulis et al., 2019).

Three datasets were collected: the ﬁrst two

with ten repetitions and the latter with two repeti-

tions of nine movements, including rest, thumb ﬂex-

ion/extension and abduction/adduction, ﬁnger ﬂex-

ion/extension (index, middle, and a combination of

ring and pinky ﬁngers), and three types of functional

grasps: pointing index, cylindrical, lateral, and tripod

grip. All participants were asked to perform bilateral

mirrored movements (with both arms). These move-

ments are shown in ﬁgure 3.

2.2 LibEMG

For data processing, feature extraction methods, clas-

siﬁer testing, and evaluation, a Python library called

LibEMG was used. This library, created by (Eddy

et al., 2023), aims to provide a simple and robust API

for developing sEMG-based projects online and of-

ﬂine.

The library offers several resources to streamline

the processes of signal reading and processing, ﬁlter-

ing, feature extraction, and selection, using various

methods commonly used in sEMG research, pattern

classiﬁcation, and model result evaluation.

In this case, the built-in function of the library was

used to organize the datasets. The function separates

signals by their classes and repetitions, generating a

set of 90 ﬁles that were split into training and testing

datasets in an 80/20 ratio.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

908

Figure 3: Acquisition protocol (Krasoulis et al., 2019).

2.3 Feature Extraction

Feature extraction aims to extract meaningful infor-

mation from the sEMG signal and is used later by the

classiﬁers (Spiewak et al., 2018). When dealing with

sEMG signals, owing to their time-based representa-

tion, feature-extraction methods commonly focus on

both the frequency and amplitude of the signal (Has-

san et al., 2020; Negi et al., 2016). Based on this,

LibEMG offers several methods commonly used in

the literature for sEMG signal analysis that can be ap-

plied individually or in groups. Twelve methods were

employed in this study.

• Discrete Fourier Transform Representation

(DFTR): Computes the energy within 6 frequency

bins of the sEMG power spectrum (Eddy et al.,

2024):

DFT R

bin

∑

i∈bin

(1)

• Root Mean Square (RMS): One of the most pop-

ular used in sEMG signal analysis, models the sig-

nal as an amplitude modulated Gaussian random

process that relates to constant force and non-

fatiguing contraction (Eddy et al., 2024; Phiny-

omark et al., 2012):

RMS =

∑

i=1

(2)

• Mean Absolute Value (MAV): Another popular

feature in sEMG signal analysis, represents the

average of the absolute values, providing a simple

measure of signal intensity over a speciﬁed time

interval (Eddy et al., 2024; Phinyomark et al.,

2012):

MAV =

∑

i=1

| (3)

• Slope Sign Changes (SSC): Represent frequency

information of the sEMG signal. Can be deﬁned

as the number of times that slope of the sEMG sig-

nal changes signs (Eddy et al., 2024; Phinyomark

et al., 2012):

SSC =

N−1

∑

i=2

f [(x

− x

i−1

) · (x

− x

i+1

)]; (4)

f (x) =

(

1, if x ≥ threshold

0, otherwise

(5)

• Integrated Absolute Value (IAV): The integral

of the absolute value (Eddy et al., 2024; Phiny-

omark et al., 2012):

IAV =

∑

i=1

| (6)

• Log Detector (LD): Provides an estimate of

the muscle contraction force (Eddy et al., 2024;

Phinyomark et al., 2012):

LD = exp

∑

i=1

log(|x

(7)

• Waveform Length (WL): Measures the com-

plexity of the sEMG signal. Deﬁned as the cu-

mulative length of the sEMG waveform over time

(Eddy et al., 2024; Phinyomark et al., 2012):

WL =

N−1

∑

i=1

i+1

− x

| (8)

• Autoregressive Coefﬁcients (AR): Models the

sEMG signal as a linear combination of its previ-

ous values, capturing temporal dependencies and

serving as a predictive feature (Eddy et al., 2024;

Phinyomark et al., 2012):

∑

p=1

i−p

+ w

(9)

• Cepstral Coefﬁcient (CC): Emphasizing peri-

odic patterns in the frequency domain by apply-

ing the inverse Fourier transform to the logarithm

Impact of Feature Extraction Optimization on Machine Learning Models for sEMG-Based Prosthesis Control

909

of the power spectrum (Eddy et al., 2024; Phiny-

omark et al., 2012):

= −a

; (10)

= −a

−

p−1

∑

l=1



1 −



p−l

(11)

• Spectral Moment (SM): A spectral moment

computed from the frequency domain. P is the

power spectrum of the signal, and f is the fre-

quencies associated with every sample of the

power spectrum (Eddy et al., 2024; Phinyomark

et al., 2012):

∑

j=1

(12)

• Wavelet Energy (WENG): Time-frequency fea-

ture that extracts the number of possible wavelets

based on a given sampling frequency argument,

then calculates the average energy of a window

for each decomposition level. Given W

is the de-

composition for level j (Eddy et al., 2024):

WENG

∑

(13)

• Difference Absolute Standard Deviation Value

(DASDV): Standard deviation of the wavelength

(Eddy et al., 2024; Phinyomark et al., 2012):

DASDV =

N − 1

N−1

∑

i=1

i+1

− x

)

(14)

2.4 Classiﬁers

After feature extraction, the data were used to train

the classiﬁcation model. In this study, tests were con-

ducted using seven classiﬁcation models based on the

reviewed literature (Eddy et al., 2024). The models

used were those deﬁned below, set to the default con-

ﬁgurations of the scikit-learn library in Python, with

no modiﬁcations to their hyperparameters or other op-

timizations.

• Linear Discriminant Analysis (LDA): Uses the

same covariance for all classes and assumes that

the data follows a normal distribution.

• k-Nearest Neighbors (KNN): Discriminates be-

tween inputs using the K closest samples in the

feature space. The LibEMG defaults to k = 5.

• Random Forest (RF): Utilizes a combination of

decision trees to differentiate between inputs.

• Quadratic Discriminant Analysis (QDA): A

quadratic classiﬁer that uses covariances speciﬁc

to each class, assuming that classes are normally

distributed.

• Naive Bayes (NB): Assumes independence

among all input features while assuming that

classes are normally distributed.

• Multilayer Perceptron (MLP): Employs

human-like ”neurons” to model data, aiding in

the discrimination between inputs.

• Support Vector Machine (SVM): Uses a hyper-

plane to maximize the distance between classes,

serving as the decision boundary for classiﬁca-

tion.

2.5 Experimental Protocol

LibEMG allows data separation into training and test-

ing sets using classes, repetitions, or subjects as pa-

rameters. The experiments were performed for each

subject using 5 k-fold cross-validation based on rep-

etitions. Two repetitions from each movement class

were randomly selected for testing, while the remain-

ing were used for model training.

An initial experiment was conducted using 5 com-

monly used time-domain feature extraction methods

(RMS, MAV, SSC, WL, and SM) across the six clas-

siﬁers to obtain a preliminary understanding of the

performance of the models. This approach aimed to

assess the effectiveness of various methods without

any further preprocessing ﬁltering, as well as to an-

alyze how these methods inﬂuence each model’s re-

sults to identify the best classiﬁer for subsequent ex-

periments.

Key classiﬁcation metrics were considered to

comprehensively evaluate the model. In this case, the

accuracy, error rate, and system instability were used.

Following the initial experiment, two additional

experiments were conducted. Each of the 12 features

was applied with the best classiﬁer obtained previ-

ously, in terms of accuracy and stability, and ﬁnally, it

was selected those features that achieved an F1-score

higher than 60% to apply again with all the classiﬁers.

For both experiments, the common classiﬁcation

metrics found in the literature were applied: accu-

racy, precision, recall, and F1-score. The equations

used are presented below, where TP represents true

positives, TN represents true negatives, FP represents

false positives, and FN represents false negatives:

• Classiﬁcation Accuracy (CA): It is the percent-

age of correctly predicted samples.

CA =

∑

i=1

ˆy

== y

(15)

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

910

where N is the total number of data

frames/predictions, ˆy

is the predicted class

label for frame i, and y

is the true class label for

frame i.

• Precision: It is the proportion of TP relative to all

examples classiﬁed as positive (TP + FP).

Precision =

T P

T P + FP

(16)

• Recall: It is the proportion of TP relative to all

positive examples (TP + FN).

Recall =

T P

T P + FN

(17)

• F1-Score: It is the harmonic mean of precision

and recall, useful when a single metric that con-

siders both precision and recall is desired.

F1-score = 2 ×

Precision × Recall

Precision + Recall

(18)

Additionally, the error and instability of each clas-

siﬁer were also obtained. The active error is the per-

centage of incorrect predictions, ignoring “No Move-

ment” class predictions. This is valuable as the No

Movement class typically correlates to “do nothing”

functionality. The instability of a classiﬁer is calcu-

lated by the number of times subsequent predictions

differ, normalized by the total number of predictions.

3 RESULTS

The results for experiment 1, using the ﬁve time-

domain features with each classiﬁer, are presented in

Tables 1 and 2. In this initial test, RF and LDA were

the classiﬁers with the best results across the three

evaluation parameters. For the S12 data, mean ac-

curacies reached 76.39% and 73.38% (standard de-

viation - 3.66 and 4.33), and mean instabilities of

4.84 and 4.30, respectively. For the S11 data, in gen-

eral, the results were worse. The RF reached the best

mean accuracy of 73.29% (standard deviation - 12.82)

with instabilities of 6.91. For both subjects, SVM and

KNN showed the worst results, with error rates above

50% and high instability.

The results obtained with each feature using RF,

the best classiﬁer in experiment 1, are presented in

Tables 3 and 4. For the S12 data, almost all features

achieved F1 scores greater than 60%, except for SSC,

which had a signiﬁcantly lower result of 51.16%. In

contrast, for the S11 data, only half of the features

achieved at least 60% F1 scores.

Finally, to assess the impact of selecting relevant

features on the ﬁnal model, the initial experiment with

the classiﬁers was redone. This time, the new features

with F1-scores above 60% were added. The results of

the ﬁnal experiment are presented in Tables 5 and 6.

For the S12 data, the ﬁnal tests improved the RF

and LDA accuracies. Despite a slight increase in in-

stability, RF remained the best overall model, achiev-

ing a mean accuracy of 79.18% (standard deviation

5.38) and instability of 5.24%. In contrast, for the S11

data, adding more features worsened the performance

of the RF despite it remaining the best overall model,

with 73.29% accuracy with a higher instability. The

most signiﬁcant impact was observed in KNN, SVM,

and MLP cases.

Table 1: Classiﬁcation results using ﬁve features (RMS,

MAV, SSC, WL, and SM) for the S12 data.

Classiﬁer CA Active error Instability

RF 76.39 24.60 4.84

LDA 73.38 28.30 4.30

QDA 68.76 29.45 6.60

NB 55.86 43.97 4.25

MLP 55.07 46.51 13.18

SVM 50.29 50.64 14.16

KNN 43.79 56.94 28.87

Table 2: Classiﬁcation results using ﬁve features (RMS,

MAV, SSC, WL, and SM) for the S11 data.

Classiﬁer CA Active error Instability

RF 73.29 26.72 6.91

LDA 63.14 35.29 7.32

QDA 61.08 37.66 10.71

MLP 44.98 54.07 16.48

NB 42.28 51.34 6.32

SVM 41.93 57.34 14.78

KNN 38.31 60.85 31.14

Table 3: Performance metrics by each feature extraction

method with the S12 data.

Feature CA Recall Precision F1

DFTR 77.33 77.33 78.22 76.82

RMS 75.43 75.43 76.37 74.99

WENG 75.16 75.16 75.96 74.57

IAV 74.56 74.56 75.42 73.91

DASDV 74.41 74.41 75.41 73.85

MAV 73.48 73.48 75.07 72.82

WL 73.23 73.23 73.93 72.47

SM 73.17 73.17 74.36 72.63

CC 72.04 72.04 73.12 71.82

AR 71.58 71.58 72.77 71.37

LD 61.15 61.15 71.88 60.23

SSC 51.85 51.85 52.10 51.16

Impact of Feature Extraction Optimization on Machine Learning Models for sEMG-Based Prosthesis Control

911

Table 4: Performance metrics by each feature extraction

method with the S11 data.

Feature CA Recall Precision F1

DFTR 73.43 73.43 74.22 73.36

WL 72.20 72.20 73.03 72.12

WENG 72.20 72.20 72.92 72.15

IAV 71.35 71.35 72.09 71.22

CC 60.32 60.32 61.12 60.19

AR 60.14 60.14 60.91 60.02

RMS 57.96 57.96 65.33 59.16

SM 54.24 54.24 64.56 59.16

DASDV 49.93 49.93 63.88 51.74

SSC 46.27 46.27 46.90 46.12

MAV 42.12 42.12 66.89 41.94

LD - - - -

Table 5: Classiﬁcation results using 10 features (DFTR,

RMS, WENG, IAV, DASDV, MAV, WL, SM, CC, AR, and

LD) with the S12 data.

Classiﬁer CA Active error Instability

RF 79.18 21.77 5.24

LDA 78.68 22.56 6.65

MLP 76.01 24.68 11.26

SVM 75.58 25.25 12.89

KNN 69.55 31.53 23.78

QDA 67.70 33.56 9.36

NB 55.90 45.06 7.35

Table 6: Classiﬁcation results using 6 features (DFTR, WL,

WENG, IAV, CC, AR) with the S11 data.

Classiﬁer CA Active error Instability

RF 72.45 27.43 8.97

MLP 69.54 30.81 16.03

LDA 66.91 33.09 12.72

SVM 66.65 33.82 19.21

KNN 64.85 33.80 32.24

NB 44.85 51.46 11.64

QDA 40.38 59.99 36.41

4 DISCUSSIONS

An important aspect noted in the literature is the need

to use amputee data. Individuals with intact limbs

generate sEMG signals based on isotonic contraction,

making it easier for the classiﬁers to recognize differ-

ent movement patterns. Instead, amputees perform

isometric contractions due to the lack of the limb.

Thus, it is fundamental to use amputee data to develop

such systems. Furthermore, their experience using

myoelectric prostheses allows for a more consistent

sEMG pattern generation. In addition, as the num-

ber of movement classes to be recognized increases,

it becomes more challenging for the model to achieve

high accuracy, particularly when classes involve sim-

ilar movements, especially for amputees. The liter-

ature shows signiﬁcant variations in the number of

movements and accuracies. However, considering the

increased complexity due to the number of classes,

accuracies ranged from 79% to 87.8%, using time-

domain features with LDA and MLP (Li and Kuiken,

2009; Li et al., 2010; Tenore et al., 2009).

The results of this work agree with the literature,

showing the relevance of working with amputee data

and the impact of experience in using myoelectric

prostheses. However, as demonstrated by comparing

the accuracies and instabilities of the classiﬁers in ex-

periments 1 and 3, using a set of features capable of

capturing valuable data patterns for the proper classi-

ﬁer can make this impact smaller.

The presented study achieved, for nine movement

classes, considering the data of the amputee with ex-

perience in the use of myoelectric prostheses (S12),

a ﬁnal mean accuracy of 79.18% (standard deviation

5.38) using RF and 78.68% (standard deviation 5.54)

using LDA with eleven time and frequency features.

In contrast, the best result achieved, using the data of

the amputee with no experience in using myoelectric

prostheses (S11), was 73.29% accuracy (standard de-

viation 3.58), also using RF; however, with the ﬁve

time-domain features.

The results showed that each feature captures

speciﬁc and different patterns in the data and that

a combination of these patterns impacts each clas-

siﬁer differently due to how each one deals with

them. Using time and frequency-domain features in

the ﬁnal experiment signiﬁcantly impacted the KNN,

SVM, and MLP, which initially had the worst over-

all results. Considering the S12 data, increases of

58.83%, 50.29%, and 38.02% in accuracies, respec-

tively, made the SVM and MLP reach accuracy val-

ues closer to the best classiﬁer performances. The

impact using the S11 data was higher for these clas-

siﬁers (69.28%, 58.96%, and 54.61%, respectively).

However, using the same data, the addition had a neg-

ative impact in RF, and in QDA using the data of both

subjects, although individually, each feature resulted

in an accuracy higher than 60%.

It is important to highlight that this study did not

include any ﬁltering or normalization steps in the data

preprocessing nor optimization of classiﬁer hyperpa-

rameters that can improve the performance of the

classiﬁcation model necessary for successful prosthe-

sis control by amputees.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

912

5 CONCLUSIONS

The results demonstrate that optimizing the feature

set with a proper classiﬁer can signiﬁcantly impact

sEMG pattern recognition performance. Although the

RF had achieved the best performance in this study,

with the best mean accuracy of 79.18% using a set of

eleven features, considering the data of the amputee

with experience in the use of myoelectric prostheses,

and 73.29% using a set of six features, considering

the data of the amputee with no experience, the most

affected models by feature optimization were KNN,

MLP, and SVM, with accuracy improvements up to

69.28%.

A possible direction for future work would be to

explore ﬁltering and normalization steps in the data

preprocessing, and deep learning classiﬁcation mod-

els aim to improve performance.

ACKNOWLEDGEMENT

This study was ﬁnanced in part by the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior –

Brasil (CAPES) – Finance Code 001. The authors

also thank FEI for their support.

REFERENCES

Andrade, N. A. (2007). Desenvolvimento de um sis-

tema de aquisic¸

ao e processamento de sinais eletro-

miogr

aﬁcos de superf

ıcie para a utilizac¸

ao no controle

de pr

oteses motoras ativas. Master’s thesis, Universi-

dade de Bras

ılia.

Castellini, C., Gruppioni, E., Davalli, A., and Sandini, G.

(2009). Fine detection of grasp force and posture by

amputees via surface electromyography. Journal of

Physiology-Paris, 103(3):255–262.

Eddy, E., Campbell, E., Phinyomark, A., Bateman, S., and

Scheme, E. (2023). Libemg: An open source library to

facilitate the exploration of myoelectric control. IEEE

Access, 11:87380–87397.

Eddy, E., Campbell, E., Phinyomark, A., Bateman, S., and

Scheme, E. (2024). Libemg documentation: Feature

extraction. Accessed: 2024-09-21.

Hassan, H. F., Abou-Loukh, S. J., and Ibraheem, I. K.

(2020). Teleoperated robotic arm movement using

electromyography signal with wearable myo arm-

band. Journal of King Saud University - Engineering

Sciences, 32(6):378–387.

Krasoulis, A., Vijayakumar, S., and Nazarpour, K. (2019).

Effect of user practice on prosthetic ﬁnger control with

an intuitive myoelectric decoder. Frontiers in Neuro-

science, 13.

Li, G. and Kuiken, T. A. (2009). Emg pattern recogni-

tion control of multifunctional prostheses by transra-

dial amputees. In 2009 Annual International Confer-

ence of the IEEE Engineering in Medicine and Biol-

ogy Society, pages 6914–6917.

Li, G., Schultz, A. E., and Kuiken, T. A. (2010). Quantify-

ing pattern recognition—based myoelectric control of

multifunctional transradial prostheses. IEEE Transac-

tions on Neural Systems and Rehabilitation Engineer-

ing, 18(2):185–192.

Negi, S., Kumar, Y., and Mishra, V. M. (2016). Feature

extraction and classiﬁcation for emg signals using lin-

ear discriminant analysis. In 2nd International Con-

ference on Advances in Computing, Communication,

Automation (ICACCA) (Fall), pages 1–6. IEEE.

NINAPRO multimodal database (2024).

https://ninapro.hevs.ch/index.html.

Phinyomark, A., Phukpattaranont, P., and Limsakul, C.

(2012). Feature reduction and selection for emg sig-

nal classiﬁcation. Expert Systems with Applications,

39(8):7420–7431.

Song, T., Yan, Z., Guo, S., Li, Y., Li, X., and Xi, F. (2023).

Review of semg for robot control: Techniques and ap-

plications. Applied Sciences, 13(17).

Spiewak, C., Islam, M., Zaman, M. A.-U., and Rahman,

M. H. (2018). A comprehensive study on emg fea-

ture extraction and classiﬁers. Open Access Journal

of Biomedical Engineering and Biosciences, pages 1–

10.

Sultana, A., Ahmed, F., and Alam, M. S. (2023). A system-

atic review on surface electromyography-based classi-

ﬁcation system for identifying hand and ﬁnger move-

ments. Healthcare Analytics, 3:100126.

Tenore, F. V., Ramos, A., Fahmy, A., Acharya, S., Etienne-

Cummings, R., and Thakor, N. V. (2009). Decoding

of individuated ﬁnger movements using surface elec-

tromyography. IEEE Transactions on Biomedical En-

gineering, 56(5):1427–1434.

Impact of Feature Extraction Optimization on Machine Learning Models for sEMG-Based Prosthesis Control

913