Prototype Reduction on sEMG Data for Instance-based Gesture

Learning towards Real-time Prosthetic Control

Tim Sziburis

1 a

, Markus Nowak

2 b

and Davide Brunelli

3 c

Department of Computer Science, Ruhr West University of Applied Sciences, 45407 M

ulheim an der Ruhr, Germany

Institute of Robotics and Mechatronics, German Aerospace Center (DLR), 82234 Weßling, Germany

Department of Industrial Engineering, DII, University of Trento, 38123 Trento, Italy

Keywords:

Surface Electromyography, Embedded Systems, Wearable Systems, Prototype Reduction, Dataset Reduction,

Instance-based Learning, Gesture Recognition, Machine Learning, Prosthetics.

Abstract:

Current systems of electromyographic prostheses are controlled by machine learning techniques for gesture

detection. Instance-based learning showed promising results concerning classiﬁcation accuracy and robustness

without explicit model training. However, it suffers from high computational demands in the prediction phase,

which can be problematic in real-time scenarios. This paper aims at combining such learning schemes with

the concept of prototype reduction to decrease the amount of data processed in each prediction step. First, a

suitability assessment of state-of-research reduction algorithms is conducted. This is followed by a practical

feasibility analysis of the approach. For this purpose, several datasets of signal classes from exerting speciﬁc

gestures are captured with an eight-channel EMG armband. Based on the recorded data, prototype reduction

algorithms are comparatively applied. The dataset reduction is characterized by the time needed for reduction

as well as the possible data reduction rate. The classiﬁcation accuracy when using the reduced set in cross-

validation is analyzed with an exemplary kNN classiﬁer. While showing promising values in reduction time

as well as excellent classiﬁcation accuracy, a reduction rate of over 99% can be achieved in all tested gesture

conﬁgurations. The reduction algorithms LVQ3 and DSM turn out to be particularly convenient.

1 MOTIVATION AND RELATED

WORK

The k nearest neighbour classiﬁcation technique

(kNN) as an exemplary instance-based machine learn-

ing scheme has shown several advantageous proper-

ties in the context of gesture recognition (intent de-

tection) based on surface electromyography (sEMG)

signals for prosthetics. In preliminary studies, kNN

showed promising results in classiﬁcation accuracy,

generalizability, as well as user study success rate

(Cipriani et al., 2011; Geethanjali, 2015; Tello et al.,

2013; Khushaba et al., 2016; Sziburis et al., 2020).

Moreover, it turned out to perform well in terms of

robustness, i. a., against sampling frequency variation

(Chen et al., 2017) and electrode shift (Li et al., 2016).

It is characterized by a comparably low implementa-

tion complexity. Furthermore, instance-based learn-

https://orcid.org/0000-0002-7741-1276

https://orcid.org/0000-0002-0840-5155

https://orcid.org/0000-0001-5110-6823

ing schemes provide the beneﬁts of no explicit math-

ematical model generation, and incrementality, i. e.

the possibility to extend the dataset by new samples

at any time with them being equally considered.

While these characteristics speak in favour of em-

bedded applicability in the context of wearable real-

time systems, an important drawback of instance-

based learning schemes is the necessity of comparing

new arriving data instances (samples) in the predic-

tion phase to all already stored ones. The needed it-

erations over all samples lead to potentially computa-

tionally intense operations, depending on the amount

of data, i. e. the samples to be iterated in every pre-

diction step (Sziburis et al., 2020).

For this reason, this work analyzes the suitability

of concepts to reduce the computational effort during

prediction in instance-based learning schemes. Al-

though no explicit model training takes place, the data

stored in memory is referred to as training data in this

paper.

Two main approaches to improve the performance

in this regard can be pointed out. The ﬁrst possibility

Sziburis, T., Nowak, M. and Brunelli, D.

Prototype Reduction on sEMG Data for Instance-based Gesture Learning towards Real-time Prosthetic Control.

DOI: 10.5220/0010327502990305

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 299-305

ISBN: 978-989-758-490-9

299

is the utilization of memory-efﬁcient, optimized data

structures, e. g. “ball-tree data structures, hashing”

(Kusner et al., 2014).

The second approach is data reduction and aims at

decreasing the amount of signal data gathered during

training and stored in memory. It can be principally

applied in a horizontal (feature space) and in a verti-

cal dimension (sample space). Aside from that, there

are techniques using approximations, such as Large

Margin Nearest Neighbour (Kusner et al., 2014) for

kNN.

Horizontal data reduction is limited by the high

variance of EMG signals. Nevertheless, the con-

cept of feature selection (or horizontal thinning) has

been applied in the context of pattern-recognition-

based prosthetic control for datasets of high feature

space dimensions, e. g. in the form of biologically in-

spired methods such as genetic algorithms and parti-

cle swarm optimization (Purushothaman, 2016). Fur-

ther concepts of horizontal data reduction are fea-

ture discretization, projection and positioning (Kus-

ner et al., 2014). These come along with dimensional-

ity reduction algorithms, e. g. PCA (G

uler and Koc¸er,

2005) and adaptions (Nagata et al., 2005) as well as

variants of LDA (Negi et al., 2016).

The concept of vertical data reduction is mainly

referred to as prototype reduction. This paper will

examine techniques of this group in the context of

EMG gesture data. Prototype stands for data instance

or sample. However, it also indicates that it refers to

speciﬁc instances that represent a larger amount of in-

stances to a certain extent.

Garc

ıa et al. (Garc

ıa et al., 2012) and Triguero et

al. (Triguero et al., 2012) provide an overview of a

variety of prototype reduction algorithms proposed.

The methods are divided into two groups: On the

one hand prototype selection (selecting a subset of in-

stances from the existing ones stored in memory, also

called vertical thinning), and on the other hand proto-

type generation (creating new instances based on the

existing ones to represent the whole dataset).

2 CONCEPTUAL APPROACH

The presented approach consists of reducing the com-

putational effort of prediction steps by decreasing the

number of samples within the gesture dataset. For this

purpose, the term reduction rate describes the num-

ber of samples ﬁnally stored in memory relative to

the number of originally recorded samples.

2.1 Theoretical Requirements

First, an assessment of the variety of algorithms re-

viewed in (Garc

ıa et al., 2012) for prototype selec-

tion and (Triguero et al., 2012) for prototype gener-

ation will be conducted, regarding the suitability of

each method. Suitable methods should not inﬂuence

the incrementality of the applied instance-based clas-

siﬁer and provide a possibility to specify the number

of prototypes in the ﬁnal set or accordingly the reduc-

tion rate beforehand. The latter characteristic will be

called size determinism in the following.

Moreover, there is the requirement of real-time ca-

pability which refers to the prediction phase of ges-

ture recognition, since this phase takes place online

and continuously decides on user satisfaction. The

real-time property is guided by timing determinism,

i. e. executing an identical number of operations per

prediction. This can be achieved by establishing the

same number of stored instances to be considered in

each prediction step. To guarantee that, size determin-

ism is inherently necessary. In order to provide fast

reaction times, the number of stored samples should

be as low as possible. Additionally, by providing size

determinism, deterministic memory demands are fa-

cilitated.

In the training phase, in contrast, sample captur-

ing and ofﬂine calculations take place, which are not

meant to be applied in real-time. Therefore, it is not

particularly needed to cope with real-time require-

ments in this phase. However, training computations

should progress as fast as possible to avoid delays

for the user between training and utilizing the gesture

prediction system.

2.2 Practical Methods

After their theoretical assessment, promising algo-

rithms will be practically evaluated in experiments on

datasets of the linear envelope of rectiﬁed EMG sig-

nals. They consist of samples captured when exerting

multiple sets of varying gesture conﬁgurations with

an eight-channel state-of-the-art armband positioned

on the forearm. With that, one sample is composed of

eight 32-bit ﬂoating-point values.

Within one dataset, four repetition blocks are al-

ternatingly recorded for each of the gestures, whereat

each repetition block consists of 400 samples (i. e.

two seconds ofﬂine data sample recording at a cap-

turing rate of 200 Hz) per gesture. Overall, three such

datasets with slightly differing sensor positionings are

recorded per gesture conﬁguration.

The following differing gesture conﬁgurations are

selected in order to test the reduction algorithms’ po-

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

300

tential dependence on inter-class properties:

• rest state, power grasp, wrist ﬂexion, wrist exten-

sion (6400 samples per dataset),

• rest state, power grasp, pointing, wrist ﬂexion,

wrist extension (8000 samples per dataset),

• rest state, power grasp, wrist ﬂexion, wrist exten-

sion, wrist pronation, wrist supination (9600 sam-

ples per dataset), and

• rest state, power grasp, pointing, wrist ﬂexion,

wrist extension, wrist pronation, wrist supination

(11200 samples per dataset).

The mentioned inter-class property dependence

could occur in the case of by-trend overlapping

classes, as in power grasp and pointing gesture where

similar groups of muscles are addressed for gesture

exertion (apart from the index ﬁnger degree of free-

dom speciﬁcally needed for the pointing gesture).

Another potential case could be wrist ﬂexion and

supination, as well as wrist extension and pronation,

which respectively address the same groups of mus-

cles, again with the exception of individual single

degrees of freedom. Another reason for selecting

several gesture conﬁguration sets is the aim to ana-

lyze the dependence of dataset reduction on the initial

dataset size due to differing sample numbers.

The achieved accuracies using the reduced sets

will be assessed in cross-validation with an exemplary

kNN classiﬁer. Since the assumption of independent

and identically distributed random variables does not

hold for samples within the same repetition block of

one gesture due to the time-dependent capturing pro-

cess, a group-wise cross-validation scheme is applied.

This means that prototype reduction takes place on

the combined set of one repetition block per gesture

(group) and the resulting prototype samples are only

validated against samples from other groups, i. e. not

against those from the same group. The average will

be taken for all possible groupings per algorithm and

gesture conﬁguration (folds).

The kNN conﬁguration is chosen to apply a

weighting of the distance d by

with a Euclidean

distance metric. With selecting k = 1, the algorithmic

runtime behaviour of the validation classiﬁer can be

reduced from O(n · log(n)) (due to sorting) to O(n)

(minimum search) with n being the number of sam-

ples in the reduced set. This was chosen as classiﬁer

setting.

Besides accuracy, the averaged runtime will be

analyzed. The runtime consists of two components,

namely the reduction time of the algorithm plus the

time needed for validation. Since the validation is

the same process for each cross-validation fold within

datasets of the same gesture conﬁguration, the valida-

tion time can be disregarded for the purpose of com-

parison. Considering that, the runtime constitutes a

representative measure for the algorithms’ reduction

times.

3 THEORETICAL ASSESSMENT

Among the prototype selection algorithms in (Garc

ıa

et al., 2012) only Random Mutation Hill Climbing

RMHC (Skalak, 1994) inherently possesses the char-

acteristic of size determinism, as it is the only method

with ﬁxed reduction. However, RMHC is a wrapper

method, which means that in each step the decision

if to select a prototype or not, a complete classiﬁer

evaluation has to take place so that high computa-

tion times during the reduction process have to be

expected. In (Garc

ıa et al., 2012, p. 425–427) it is

shown that this assumption holds in real use-cases for

variably sized sets of data. Exemplary tests on EMG

datasets conﬁrmed this behaviour so that RMHC was

excluded from further consideration.

Besides the ﬁxed reduction prototype selection al-

gorithms, there may be also mixed reduction methods

which provide the property of determinism with re-

spect to the number of samples contained in the ﬁnal

training set. Nevertheless, the algorithms of that cat-

egory described in (Garc

ıa et al., 2012) are all wrap-

per methods, too. Due to the respective high execu-

tion times as mentioned before, these methods are not

considered within the scope of this work.

In terms of prototype generation, there is a variety

of ﬁxed reduction approaches, structured as follows:

• Positioning Adjustment, Condensation: Vector

Quantization VQ (Qiaobing Xie et al., 1993),

Learning Vector Quantization (LVQ) methods:

LVQ3 (Kohonen, 1990), LVQ with Training

Counter LVQTC (Odorico, 1997), Decision Sur-

face Mapping DSM (Geva and Sitte, 1991),

• Positioning Adjustment, Hybrid Approach: Parti-

cle Swarm Optimization PSO (Nanni and Lumini,

2009),

• Centroid-based Condensation: Bootstrap Tech-

niques BTS3 (Hamamoto et al., 1997), Adap-

tive Condensing Algorithm Based on Mixtures of

Gaussians MGauss (Lozano et al., 2006),

• Space-splitting: Chen Algorithm (Chen and

zwik, 1996).

LVQTC does not provide size determinism so that

it was not taken into account for further evaluation,

while the other algorithms will be practically exam-

ined.

Prototype Reduction on sEMG Data for Instance-based Gesture Learning towards Real-time Prosthetic Control

301

Besides the ﬁxed algorithms, also in prototype

generation there are mixed methods again, which may

principally also provide size determinism:

• Positioning Adjustment, Condensation: Gradient

Descent and Deterministic Annealing MSE (De-

caestecker, 1997), Hybrid LVQ3 HYB, (Kim and

Oommen, 2003), LVQ with Pruning LVQPRU (Li

et al., 2005),

• Positioning Adjustment, Hybrid Approach: Pro-

totype Selection Clonal Selection Algorithm

PSCSA (Garain, 2008) using an artiﬁcial im-

mune system model, Evolutionary Nearest Pro-

totype Classiﬁer ENPC (Fern

andez and Isasi,

2004), Adaptive Michigan PSO AMPSO (Cer-

vantes et al., 2007),

• Centroid-based Hybrid Approach: Integrated

Concept Prototype Learner ICPL2 (Lam et al.,

2002).

In turn, some of these are wrapper methods

(ENPC, AMPSO) and hence not considered with re-

gard to the previously mentioned reason. The ﬁlter

and semi-wrapper methods MSE, HYB, and ICPL2

do not provide size determinism.

Overall, the remaining methods to practically an-

alyze in the subsequent section are MGauss, BTS3,

Chen, LVQ3, LVQPRU, DSM, VQ, and PSCSA.

4 PRACTICAL EVALUATION

Following the experimental conﬁguration presented

in section 2.2, exemplary experiments showed that

a reduction to 20 prototypes in the ﬁnal set yields

a promising performance (regarding reduction time

and classiﬁcation accuracy), which was hence set as

the goal.

Figures 1 to 4 illustrate the results of cross-

validation accuracies and runtimes per algorithm and

gesture conﬁguration, averaged over all validation

folds for the corresponding datasets.

The time measurements refer to the usage of the

open-source (GPLv3) tool KEEL (Knowledge Ex-

traction based on Evolutionary Learning) (Triguero

et al., 2017) on GNU/Linux kernel 4.18.0-25 (Intel

Core i7-8550U, 1.80 GHz, 7869 MiB RAM) with

only essential system processes running.

Each repetition block is supposed to have the

same size, just slightly differing due to sampling

rate deviations when the training data was cap-

tured. The training data is reduced to 20 samples

(i. e. 20 times eight channel values with a 32-bit

ﬂoating-point number per channel) in each procedure.

MGauss BTS3 Chen LVQ3 -PRU DSM VQ PSCSA

Reduction Method

100

Accuracy [%]

Median

25%-75%

9%-91%

(a) Accuracy

MGauss BTS3 Chen LVQ3 -PRU DSM VQ PSCSA

Reduction Method

Runtime [ms]

Median

25%-75%

9%-91%

(b) Reduction time

Figure 1: Averaged performance results for datasets with

four gestures (rest, power, ﬂexion, extension).

Figure 1 shows equally high classiﬁcation accu-

racies (100%) for all algorithms in all cases, with

small exceptions for BTS3 and VQ exposing some

outliers and therefore higher variance. However, even

these methods yield 100% as median and over 90% as

mean. In terms of time, PSCSA is noticeably worse

than the other algorithms which all deliver times un-

der 2 ms vs. 17 ms for PSCSA. Due to this high dis-

crepancy already in a case with only four gestures and

better performance of the other methods in all cases,

PSCSA was excluded from further consideration.

When additionally including the pointing gesture

(ﬁgure 2), i. e. introducing a by-trend overlapping

class, differences among the algorithms can be ob-

served. BTS3 and VQ again show high variances in

accuracy. In comparison to the previous dataset, they

perform even worse, reducing the accuracies from

slightly over 90% in mean to about 85% and 90%

respectively. The other algorithms’ accuracy means

are still close to or at 100%. All medians still lay at

100%. Chen, MGauss and LVQPRU have a compara-

bly broad variance of runtime, leading to the assump-

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

302

MGauss BTS3 Chen LVQ3 -PRU DSM VQ

Reduction Method

100

Accuracy [%]

Median

25%-75%

9%-91%

(a) Accuracy

MGauss BTS3 Chen LVQ3 -PRU DSM VQ

Reduction Method

0.5

1.5

Runtime [ms]

Median

25%-75%

9%-91%

(b) Reduction time

Figure 2: Averaged performance results for datasets with

ﬁve gestures (rest, power, point, ﬂexion, extension).

tion of higher nondeterminism during execution. Fur-

thermore, these three algorithms yielded the highest

means and medians in runtime.

A dataset without point gesture but with wrist

pronation and wrist supination (ﬁgure 3) does not

seem to affect the accuracy of BTS3 anymore (con-

stantly at 100%), although VQ’s results even more

(accuracy mean of 70% but still median of 100%).

The other schemes provide 100% in all cases. The

timings of Chen and LVQPRU still show the high-

est mean and median as well as variance. Among the

remaining algorithms, MGauss and VQ are slightly

slower than BTS3, LVQ3, and DSM.

Testing all gestures, i. e. including pointing ges-

ture as well as wrist rotations (ﬁgure 4), summarizes

previous observations, i. e. the reduced accuracy of

BTS3 and VQ with a mean of about 80% while the

others provide 100%. All algorithms expose a me-

dian accuracy of 100%. Regarding timing, Chen and

LVQPRU have the highest means and medians. Fur-

thermore, their timing variance can be considered as

high, leading to the presumption of reduced time de-

terminism.

MGauss BTS3 Chen LVQ3 -PRU DSM VQ

Reduction Method

100

Accuracy [%]

Median

25%-75%

9%-91%

(a) Accuracy

MGauss BTS3 Chen LVQ3 -PRU DSM VQ

Reduction Method

0.5

1.5

2.5

Runtime [ms]

Median

25%-75%

9%-91%

(b) Reduction time

Figure 3: Averaged performance results for datasets with

six gestures (rest, power, ﬂexion, extension, pronation,

supination).

5 DISCUSSION AND OUTLOOK

The results showed that the approach of prototype

reduction seems promising to cope with the main

drawback of instance-based learning in the context

of EMG-based gesture detection for real-time embed-

ded applications, such as prosthetic control. The ex-

amined algorithms are suitable to be used with any

instance-based learning technique; the general feasi-

bility could be demonstrated with a kNN classiﬁca-

tion scheme. The reduction of data processed in pre-

diction to a deterministic and low number of 20 data

samples (i. e. reduction rate of over 99%) while pre-

serving excellent classiﬁcation accuracy could be suc-

cessfully shown. With that, data with no informa-

tional content regarding classiﬁcation can be disre-

garded. The very high possible reduction rate points

out the prevalent existence of that kind of data. How-

ever, detailed analyses of the dependence on reduction

rate have to be conducted, considering a higher num-

ber and greater variety of gesture datasets as well as

Prototype Reduction on sEMG Data for Instance-based Gesture Learning towards Real-time Prosthetic Control

303

MGauss BTS3 Chen LVQ3 -PRU DSM VQ

Reduction Method

100

Accuracy [%]

Median

25%-75%

9%-91%

(a) Accuracy

MGauss BTS3 Chen LVQ3 -PRU DSM VQ

Reduction Method

Runtime [ms]

Median

25%-75%

9%-91%

(b) Reduction time

Figure 4: Averaged performance results for datasets with

seven gestures (rest, power, point, ﬂexion, extension, prona-

tion, supination).

the inﬂuence of the chosen classiﬁer and its parame-

terization.

While the main aspect of coping with compu-

tationally high demands in real-time instance-based

learning is the prediction phase, the training phase has

no explicit real-time requirement as it takes place in

an ofﬂine manner. However, the training time should

be as short as possible in order to avoid delays for

the user before being able to interact with the sys-

tem, which is especially important in actual wearable

prosthetic devices. On the present simulation sys-

tem, prompt training runtimes under 0.5 ms could be

achieved in all datasets for the best-performing algo-

rithms (including cross-validation time). The timing

characteristics have to be further evaluated on differ-

ing types of embedded systems, under consideration

of speciﬁc requirements in prosthetic applications.

For all used combinations of gestures, especially

LVQ3 and DSM (both based on Learning Vector

Quantization, LVQ) turned out to be of particular in-

terest. The PSCSA algorithm was already disregarded

in an early stage due to around 16 times higher run-

times in comparison to the other methods, which can

be related to the complexity of the artiﬁcial immune

system model behind.

While all evaluated algorithms interestingly

yielded 100% as median after reducing the origi-

nal datasets to 20 samples, LVQ3, DSM, Chen and

MGauss provided 100% classiﬁcation accuracy even

as mean value. Among these methods, MGauss and

especially Chen exhibit comparably high reduction

times (together with LVQPRU).

With a low runtime of about 0.2 ms in most cases

and a low runtime variance, LVQ3 and DSM seem to

be the most suitable among the tested reduction al-

gorithms for real-time scenarios and embedded ap-

plications. This will be subject to further research,

with speciﬁcally adapting the algorithmic program-

ming and hardware frameworks as well as an in-depth

examination of time complexity and tuning all param-

eters involved. Moreover, user studies have to be con-

ducted with participants performing target achieve-

ment tests in online gesture classiﬁcation and real-

time prosthetic control applications.

ACKNOWLEDGMENTS

This work is supported by the Ministry of Eco-

nomics, Innovation, Digitization and Energy of the

State of North Rhine-Westphalia and the European

Union, grants GE-2-2-023A (REXO) and IT-2-2-023

(VAFES).

REFERENCES

Cervantes, A., Galv

an, I., and Isasi, P. (2007). An adap-

tive michigan approach pso for nearest prototype clas-

siﬁcation. In Mira, J. and

Alvarez, J. R., editors,

Nature Inspired Problem-Solving Methods in Knowl-

edge Engineering, pages 287–296, Berlin, Heidel-

berg. Springer Berlin Heidelberg.

Chen, C. and J

zwik, A. (1996). A sample set condensation

algorithm for the class sensitive artiﬁcial neural net-

work. Pattern Recognition Letters, 17(8):819 – 823.

Chen, H., Zhang, Y., Zhang, Z., Fang, Y., Liu, H., and

Yao, C. (2017). Exploring the relation between emg

sampling frequency and hand motion recognition ac-

curacy. In 2017 IEEE International Conference on

Systems, Man, and Cybernetics (SMC), pages 1139–

1144.

Cipriani, C., Antfolk, C., Controzzi, M., Lundborg, G.,

Rosen, B., Carrozza, M. C., and Sebelius, F. (2011).

Online myoelectric control of a dexterous hand pros-

thesis by transradial amputees. IEEE Transactions

on Neural Systems and Rehabilitation Engineering,

19(3):260–270.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

304

Decaestecker, C. (1997). Finding prototypes for nearest

neighbour classiﬁcation by means of gradient descent

and deterministic annealing. Pattern Recognition,

30(2):281 – 288.

Fern

andez, F. and Isasi, P. (2004). Evolutionary design of

nearest prototype classiﬁers. Journal of Heuristics,

10(4):431–454.

Garain, U. (2008). Prototype reduction using an artiﬁcial

immune model. Pattern Anal. Appl., 11:353–363.

Garc

ıa, S., Derrac, J., Cano, J., and Herrera, F. (2012).

Prototype selection for nearest neighbor classiﬁca-

tion: Taxonomy and empirical study. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

34(3):417–435.

Geethanjali, P. (2015). Comparative study of pca in clas-

siﬁcation of multichannel emg signals. Australasian

Physical & Engineering Sciences in Medicine,

38(2):331–343.

Geva, S. and Sitte, J. (1991). Adaptive nearest neighbor pat-

tern classiﬁer. IEEE transactions on neural networks

/ a publication of the IEEE Neural Networks Council,

2:318–22.

uler, N. F. and Koc¸er, S. (2005). Classiﬁcation of emg

signals using pca and fft. Journal of Medical Systems,

29(3):241–250.

Hamamoto, Y., Uchimura, S., and Tomita, S. (1997). A

bootstrap technique for nearest neighbor classiﬁer de-

sign. IEEE Transactions on Pattern Analysis and Ma-

chine Intelligence, 19(1):73–79.

Khushaba, R. N., Al-Timemy, A., Al-Ani, A., and Al-

Jumaily, A. (2016). Myoelectric feature extraction

using temporal-spatial descriptors for multifunction

prosthetic hand control. In 2016 38th Annual In-

ternational Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC), pages 1696–

1699.

Kim, S.-W. and Oommen, B. J. (2003). A brief taxonomy

and ranking of creative prototype reduction schemes.

Pattern Analysis & Applications, 6(3):232–244.

Kohonen, T. (1990). The self-organizing map. Proceedings

of the IEEE, 78(9):1464–1480.

Kusner, M. J., Tyree, S., Weinberger, K., and Agrawal, K.

(2014). Stochastic neighbor compression. In Proceed-

ings of the 31st International Conference on Machine

Learning - Volume 32, ICML’14, pages II–622–II–

630. JMLR.org.

Lam, W., Keung, C.-K., and Liu, D. (2002). Discovering

useful concept prototypes for classiﬁcation based on

ﬁltering and abstraction. Pattern Analysis and Ma-

chine Intelligence, IEEE Transactions on, 24:1075–

1090.

Li, J., Manry, M. T., Yu, C., and Wilson, D. R.

(2005). Prototype classiﬁer design with pruning. In-

ternational Journal on Artiﬁcial Intelligence Tools,

14(01n02):261–280.

Li, Q. X., Chan, P. P. K., Zhou, D., Fang, Y., Liu, H., and

Yeung, D. S. (2016). Improving robustness against

electrode shift of semg based hand gesture recogni-

tion using online semi-supervised learning. In 2016

International Conference on Machine Learning and

Cybernetics (ICMLC), volume 1, pages 344–349.

Lozano, M., Sotoca, J. M., S

anchez, J. S., Pla, F., Pkalska,

E., and Duin, R. P. W. (2006). Experimental study

on prototype optimisation algorithms for prototype-

based classiﬁcation in vector spaces. Pattern Recogn.,

39(10):1827–1838.

Nagata, K., Adno, K., Magatani, K., and Yamada, M.

(2005). A classiﬁcation method of hand movements

using multi channel electrode. In 2005 IEEE Engi-

neering in Medicine and Biology 27th Annual Confer-

ence, pages 2375–2378.

Nanni, L. and Lumini, A. (2009). Particle swarm opti-

mization for prototype reduction. Neurocomputing,

72:1092–1097.

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

traction and classiﬁcation for emg signals using linear

discriminant analysis. In 2016 2nd International Con-

ference on Advances in Computing, Communication,

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

Odorico, R. (1997). Learning vector quantization with

training count (lvqtc). Neural networks : the ofﬁcial

journal of the International Neural Network Society,

10(6):1083—1088.

Purushothaman, G. (2016). Myoelectric control of pros-

thetic hands: State-of-the-art review. Medical De-

vices: Evidence and Research, Volume 9:247–255.

Qiaobing Xie, Laszlo, C. A., and Ward, R. K. (1993). Vec-

tor quantization technique for nonparametric classiﬁer

design. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 15(12):1326–1330.

Skalak, D. B. (1994). Prototype and feature selection by

sampling and random mutation hill climbing algo-

rithms. In Cohen, W. W. and Hirsh, H., editors, Ma-

chine Learning Proceedings 1994, pages 293 – 301.

Morgan Kaufmann, San Francisco (CA).

Sziburis, T., Nowak, M., and Brunelli, D. (2020). knn learn-

ing techniques for proportional myocontrol in pros-

thetics. In Torricelli, D., Akay, M., and Pons, J. L.,

editors, 5th International Conference on NeuroReha-

bilitation (ICNR2020) and 5th International Sympo-

sium on Wearable Robotics (WeRob2020), Converg-

ing Clinical and Engineering Research on Neuroreha-

bilitation IV. Springer.

Tello, R. M. G., Bastos-Filho, T., Costa, R. M., Frizera-

Neto, A., Arjunan, S., and Kumar, D. (2013). Towards

semg classiﬁcation based on bayesian and k-nn to con-

trol a prosthetic hand. In 2013 ISSNIP Biosignals and

Biorobotics Conference: Biosignals and Robotics for

Better and Safer Living (BRC), pages 1–6.

Triguero, I., Derrac, J., Garcia, S., and Herrera, F. (2012). A

taxonomy and experimental study on prototype gener-

ation for nearest neighbor classiﬁcation. IEEE Trans-

actions on Systems, Man, and Cybernetics, Part C

(Applications and Reviews), 42(1):86–100.

Triguero, I., Gonz

alez, S., Moyano, J., Garc

ıa, S., Alcala-

Fdez, J., Luengo, J., Fern

andez, A., Del Jesus, M. J.,

Sanchez, L., and Herrera, F. (2017). Keel 3.0: An

open source software for multi-stage analysis in data

mining. International Journal of Computational In-

telligence Systems, 10(1):1238–1249.

Prototype Reduction on sEMG Data for Instance-based Gesture Learning towards Real-time Prosthetic Control

305