Transfer-Modal Extraction of Surface EMG Features for Upper Limb

Motor Classiﬁcation

Vedant Mangrulkar and Madhav Rao

International Institute of Information Technology Bangalore, Bangalore, India

{vinit.mangrulkar, mr}@iiitb.ac.in

Keywords:

Accelerometer, EMG, Classiﬁer, ML, Features, Upper Limb, Synthetic Features, Random Forest.

Abstract:

Surface Electromyography (sEMG) signals provide critical insights into muscular activity, aiding action clas-

siﬁcation and monitoring muscular disorders. However, their reliability is hindered by noise and unstructured

data. Despite the advancements in machine learning, large datasets are essential to address these challenges

and enhance decoding accuracy for further development. Hence, this work attempts to predict the sEMG fea-

tures from the accelerometer signals in a view to generate synthetic data which is useful for further develop-

ments around this physiological signal. This work examines the correlation between accelerometer-generated

sEMG features and those from original sEMG signals for four upper limb actions wrist ﬂexion, wrist exten-

sion, wrist closing and wrist vibration focusing on the ﬂexor carpi ulnaris and extensor carpi radialis muscles.

Synthesized features are augmented with original features to train an ML model, achieving 91% accuracy

on unseen original sEMG features. This work showcases a viable solution to generate more sEMG features

corresponding to the actions under test from an altogether different modality. This work is a step towards

synthesizing EMG signals and features for human limb movements which offers a strong platform to design

imitation learning for rehabilitation systems in the future.

1 INTRODUCTION

The human hand is one of the most important part

of our body. It plays a crucial role in our daily life

helping us perform a wide range of activities start-

ing from basic tasks to complex exercises (Schreuders

et al., 2019). The wrist, a key joint in our hand move-

ment, allows for essential movements such as ﬂex-

ion, extension and rotation, which are very important

for various activities involving gripping, lifting and

many other ﬁne motor skills (Eschweiler et al., 2022).

Unfortunately, some unexpected events or accidents

leads to the partial or complete impairments in upper

limb functions. Hence the patients are prescribed with

speciﬁc upper limb physiotherapy based exercises

towards effective rehabilitation (Jonna et al., 2024;

Jonna and Rao, 2022; P et al., 2023). The recovery

process although is extremely slow, but is considered

effective owing to continuous improvement (Vinay

et al., 2021). Over the years, the researchers have

explored ways of monitoring progress and effective-

ness of the therapy sessions through dedicated care-

takers or paid physicians at home or clinics and cen-

https://orcid.org/0000-0003-2278-9148

ters. The visual evaluation is always marred by hu-

man bias and remains inconsistent across individu-

als undergoing therapy and also along the evaluators.

Various methods around sensory systems to automate

and nullify the human bias is explored in the past, but

a more simpliﬁed and robust variant which is reliable

for practising is yet to be found.

The motion classiﬁcation for human limb re-

habilitation presents an opportunity to develop au-

tomation in the personalized therapy program with

real-time feedback and assessment systems to en-

sure that the routine therapy are performed accu-

rately and effectively (Novak et al., 2012; Kwakkel

et al., 2008). Automation without much support from

care-takers helps the former to build their conﬁdence

and self-belief, besides giving much relief to the lat-

ter. Automated features aids in progress-tracking

and continuous-monitoring, effectively improving the

therapy sessions (Hiengkaew et al., 2012). Another

distinct and radical approach of employing dynami-

cally adaptive training system is introduced and fur-

ther discussed in (Loureiro and Harwin, 2007) which

works on self-tuning the therapy based on the out-

come.

Hence an automated classiﬁcation of patient

Mangrulkar, V. and Rao, M.

Transfer-Modal Extraction of Surface EMG Features for Upper Limb Motor Classiﬁcation.

DOI: 10.5220/0013157800003911

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 721-728

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

721

limb movements is therefore extremely crucial in

the scheme of rehabilitation process (Chandrasekhar

et al., 2020a).

Electromyography (EMG) signals capture the

muscular activity of the human limb, and same is re-

ported in many literature’s for recognizing the upper

limb actions (Vinay et al., 2022; Reddy et al., 2023).

More macro-level evaluation pertains to the actions

performed by the individuals under rehabilitation, and

in-conjunction with EMG signal is a useful parameter

to analysis. Hence EMG signal based upper and lower

limb rehabilitation designs are found most suitable for

evaluating the recovery process.

The invasive variant of acquiring EMG signals is

more accurate, but does not ﬁt this requirement of

continuous evaluation. Hence surface based EMG

signals also referred to as sEMG signals are tapped

at the surface of the human body. Although these sig-

nals when probed at the surface is conducive to noise,

but are processed with some success to ﬁnd the signa-

tures mapped to the limb movements (Chandrasekhar

et al., 2020b). The sEMG signals are found to carry

critical information of limb movements and hence are

aimed toward the rehabilitation setups. However the

usage of sEMG signals for the classiﬁcation of the

motor movements has its own share of problems. The

sEMG signals are marred with huge baseline noise in

addition to the distortion effects which leads to a low

SNR (Campanini et al., 2020).

The occupancy of baseline noise signal makes it

extremely difﬁcult to extract discriminative features,

while the

distortion alters the temporal and phase compo-

nents of the signal (Merletti and Cerone, 2020). The

sEMG signals are also sensitive to several other fac-

tors including human anatomy, electrode placements

and muscle fatigues (Merletti and Cerone, 2020).

Other sensory systems including motion capture

systems, IMU, Electromagnetic Tracking Sensor and

pressure sensors fall short in reliably detecting mo-

tor imagery due to several reasons (Mason and Birch,

2003; Chizari et al., 2020). Additionally, the mo-

tion capture systems are cost ineffective when com-

pared to sEMG or MPU(motion processing unit) sen-

sors, and are predominantly built on camera subsys-

tem, which invades privacy. The video grab is typi-

cally of high bandwidth and demand high computer

power to process which makes it unviable for clinical

trials (Vitali and Perkins, 2020).

The pressure sensors on the other hand are sen-

sitive to their placements and alignments, which

proves to be challenging for rehabilitation purposes

(Lawrence et al., 2014).

Overall, the sEMG based systems are found to be

an apt solution for the domestic and clinical setups

where affordability is a major factor, however build-

ing the whole system requires more reliable data to-

wards training the classiﬁer. One way to deal with this

challenge is to acquire large sEMG dataset for train-

ing, thereby making the model resilient to high degree

of variability among the individuals and also along the

acquisition factors. However, compiling sEMG sig-

nals requires comprehensive effort and hence for most

of the custom applications, sEMG signals remain un-

favourable. A synthesized set of sEMG dataset for

building classiﬁer is one of the solution towards ﬁll-

ing the aforementioned void. An electrical model

based sEMG features are reported in the past (Vinay

et al., 2022), however these remains non-reliable for

the real-world applications even with the addition of

Gaussian modeled noise component to the synthe-

sized features. A suitable alternative is to derive syn-

thetic features from a different set of physical sensors

which preferably is less sensitive to motion artifacts.

Accelerometers are found reliable for motion classi-

ﬁer both for upper and lower limb, however these

are not equipped to offer distal information instead

offers proximal information when utilized in the hu-

man body. For instance, the muscle activity for wrist

movement is captured by sEMG signal even at the dis-

tal end of the upper arm, whereas MPUs are likely to

be positioned near the wrist joint to capture reliable

signal. Consequently the proximal accelerometer sen-

sory signal carries the necessary limb movement in-

formation which can be extracted to the correspond-

ing synthetic distal sEMG signal features. These syn-

thetic features supplement to build large dataset for

realizing ML classiﬁer model to work for real-world

automated rehabilitation tasks. This paper demon-

strates the principle working of the proposed transfer

of modality for training model and evaluates the same

using original sEMG signal features and also for the

mixed (original and synthetic) dataset. The main con-

tributions of the paper are: i) Prediction of sEMG fea-

tures from the extracted real time accelerometer data,

and further validating the same with ensemble model

to achieve reliable results, and

ii) Extracted sEMG features show comparable

performance with respect to other features extracted

from original sEMG signal. The dataset and model

ﬁles are made freely available at (Mod, ) for further

usage to the researchers and designers community.

2 PROPOSED METHOD

The proposed framework as shown in the Figure 1

comprises of three primary systems. The ﬁrst system

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

722

Figure 1: Proposed framework#1 comprising of Ensemble

model to extract sEMG features from MPU signals and fur-

ther supplying to a Classiﬁer model. Here Model#1 is En-

semble model, and Model#2 is a Classiﬁer model.

focuses on data acquisition from MPU sensors.

The second system involves prediction of the

sEMG features from the ﬁltered accelerometer data.

The IMU data for the speciﬁc movements is taken to

label the data appropriately and further train the ex-

traction model. The last system engages in the classi-

ﬁcation of four identiﬁed upper limb movements us-

ing these synthetic features extracted from IMU data.

ML MODEL 1 is the prediction model which predicts

the features from the ﬁltered data and ML MODEL 2

is the classiﬁer model which classiﬁes the four mo-

tions.

2.1 Dataset

Surface EMG signals are captured using the Mus-

cle BioAmp Patchy sensor, which offers an input

impedance of 10

Ω and includes a bandpass ﬁlter

with a cutoff frequency of 72-720 Hz (Upside Down

Labs, 2024). Motion data of the wrist is measured us-

ing an Inertial Measurement Unit (IMU), speciﬁcally

the MPU 6050, which provides the accelerometer

and gyroscope data across all three axes (InvenSense,

2013). The MPU 6050 sensor and the sEMG sen-

sors are interfaced to the Arduino Nano board via I2C

communication and analog pins, respectively. The ac-

celerometer data offers the vector and movement in-

formation along the 3 axis and are preferred over gy-

roscopes which are more useful for rotational actions.

Hence in this work, only accelerometer data is em-

ployed to augment the sEMG features.

Figure 2(a and b) shows the setup to acquire the

sEMG signal from the targetted muscle group, and

IMU for the same upper limb actions. The IMU setup

was designed to

be wearable, and user-friendly, while sEMG elec-

trodes are positioned appropriately to ensure effective

data sampling during wrist movements. Two Bioamp

sEMG sensors were employed to collect the signals

from the targeted muscles at the same time, while the

IMU data is also captured for the four identiﬁed upper

limb movements. Necessary informed ethical consent

was taken from 8 healthy subjects with a declaration

Figure 2: Snapshot illustrating the setup to acquire signals

from IMU and sEMG for the deﬁned upper limb move-

ments. IMU sensors enclosed in a black band is wrapped

around the palm of the hand, whereas the sEMG signals are

acquired from the electrodes positioned at the forearm. (a)

shows the position of electrodes on the ﬂexor carpi ulnaris

muscles, and (b) shows the position of electrodes on the ex-

tensor carpi radialis muscles.

approving that the subjects had no prior weakness or

injury, nor any prior surgery performed in the past on

any part of the upper or lower limb.

These participants were evenly distributed be-

tween 4 females and 4 males to present generic re-

sults. Participants’ characteristics included an aver-

age age of 19.5 ± 0.87 years and an average weight

of 62.32 ± 16.71 kg. All participants provided writ-

ten informed consent before participating in the ex-

periments. The data was collected in accordance

with the Helsinki Declaration of 1975, as revised in

2000 (Ashcroft et al., 2008). The subjects were in-

structed to perform four distinct wrist motions: wrist

ﬂexion, wrist extension, wrist closing(extension and

contraction of ﬁngers), and wrist vibration(shaking of

hand). The dataset collected for training the model

comprised a total of 108,656 samples. Figure 3 shows

one such instance of the real-time sEMG signal along

with the corresponding accelerometer and gyroscope

data in all 3 axes for all the four aforementioned

wrist actions. To ensure data randomness, subjects

performed wrist motions in a randomized sequence.

This method mitigates potential biases and enhances

the robustness of the dataset for better analysis (Is-

mail Fawaz et al., 2018). Electrodes were positioned

on the ﬂexor carpi ulnaris and extensor carpi radialis

muscles, on the skin. Real-time signal data was cap-

tured and stored in readable ﬁles prior to feature ex-

traction from sEMG signals and further labelling the

signal for the corresponding actions.

2.2 EMG Feature Extraction from IMU

Data

The signals were uniformly collected at a sampling

rate of 500 Hz, as proven efﬁcient in the litera-

ture (Chen et al., 2017). Accelerometer data under-

went Kalman ﬁltering (Process Noise Covariance(Q)

: 0.01*eye(3), Measurement Noise Covariance(R) :

0.1*eye(3), Initial Estimation Error Covariance(P) :

Transfer-Modal Extraction of Surface EMG Features for Upper Limb Motor Classiﬁcation

723

Figure 3: Pre-processed sEMG signal with the correspond-

ing Accelerometer signals for the four upper limb wrist ac-

tions - ﬂexion, extension, closing of wrist, and vibrations.

eye(3), State Transition and Covariance : eye(3))

to signiﬁcantly enhance accuracy by reducing noise

and improving the reliability of the IMU data (Farag,

2020). Surface electromyography (sEMG) signals

were processed with the help of a Butterworth 4

order band-pass ﬁlter for enhanced efﬁciency (Shi,

2012). Following data acquisition, feature extrac-

tion was conducted for a window size of 50 sam-

ples (100 ms) with a step size of 1 sample (2 ms)

to retain higher order of dataset. Time-domain fea-

tures including Integrated EMG (iEMG), Root Mean

Square (RMS), and Variance were extracted and an-

alyzed for each motion. These features serve as the

most discriminative inputs for further motion classiﬁ-

cation tasks (Phinyomark et al., 2012).











iEMG =

|EMG(t)|dt

RMS =

∑

i=1

(EMG

)

Variance =

∑

i=1

(EMG

− µ)

(1)

2.3 Model Feature Prediction

Filtered accelerometer data and extracted sEMG fea-

tures serve as input and output, respectively, for pre-

dicting sEMG features. Each feature (iEMG, RMS,

Variance) is predicted independently. To predict

iEMG and RMS features, we employed an ensemble

model combining a Convolutional Neural Network

(CNN) and a Random Forest (RF) to enhance predic-

tive performance. CNN was employed for its abil-

ity to capture spatial hierarchies in the data through

convolutional layers (LeCun et al., 2015). Random

Forest was used for its robustness and ability to han-

dle nonlinear relationships (Khalilia et al., 2011). For

predicting Variance, an ensemble model combining

Figure 4: Original features versus trained ML Predicted fea-

tures for all four selected upper limb actions.

Random Forest (estimators=500, depth=30, MinSam-

plesSplit=3, MinSamplesLeaf=2, RandomState=42)

and LightGBM(estimators=500, LearningRate=0.05,

depth=30, leaves=40, MinchildSamples=20, Ran-

domState=42) was utilized for improved accuracy.

LightGBM is particularly effective for time series

feature prediction due to its fast training speed, low

memory usage, and ability to manage large datasets

and complex feature interactions efﬁciently (Ke et al.,

2017). The outputs from these models were com-

bined to achieve a more robust and accurate pre-

dictions. The CNN architecture consisted of mul-

tiple convolutional layers followed by pooling lay-

ers, batch-normalization, and dropout layers to pre-

vent over-ﬁtting, and dense layers for the ﬁnal output

(Krizhevsky et al., 2012; Ioffe and Szegedy, 2015).

The RF model was tuned with hyperparameters such

as the number of trees, maximum depth, and mini-

mum samples split to achieve optimal performance.

For feature engineering, rolling statistics : mean and

standard deviation are calculated with a window size

of 10 to enhance the input feature set.

The predictions from both models were then en-

sembled using a weighted average approach. The

weights were determined based on the individual per-

formance of each model on a validation set. This en-

semble approach leveraged the strengths of both mod-

els, leading to improved overall performance com-

pared to using each model individually (Zhou, 2012).

The prediction model was trained for 70% of the

dataset, 10% for validation, and remaining 20% for

testing. The ensemble model achieved R

value (co-

efﬁcient of determination) of 0.789 and 0.775 for the

prediction of iEMG and RMS features with respect to

the original extracted features respectively. The en-

semble model for the prediction of Variance showed

value of 0.715. Figure 4 showcases the plot of all

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

724

the three predicted features in contrast with the orig-

inal features. The R

value is a crucial metric in re-

gression analysis as it quantiﬁes the proportion of the

variance in the dependent variable that is predictable

from the independent variables. R

is a key indica-

tor of model performance and reliability in predictive

analytics (Frost, 2019).

2.4 Motion Classiﬁer Model

Post the prediction of the sEMG features, the data

was labelled for the classiﬁcation of the wrist move-

ments. The classiﬁer model used 70% of the dataset

for training, 10% for validation, and remaining 20%

of the dataset for testing. A grid search was con-

ducted to optimize hyperparameters, using a 5-fold

cross-validation for all models. The 70-10-20 split

for training, validating and testing dataset is proved

to be efﬁcient due to its balance between training,

and evaluation (Arlot and Celisse, 2009). Between

the various models trained, RF offered the best results

for the classiﬁcation of the wrist movements with pa-

rameters(estimators = 200, depth = 20, samplesSplit

= 5, MinSamplesLeaf = 2, RandomState = 42, Cri-

terion = ’gini’, bootstrap = True, features = ’sqrt’)

RF is widely recognized as an effective classiﬁer for

movement recognition tasks due to its robustness and

ability to handle complex, non-linear relationships

in data. It combines multiple decision trees to im-

prove predictive accuracy and reduce the risk of over-

ﬁtting, making it particularly suitable for classifying

movements from sensor data (Pal, 2005). Amongst

the other classiﬁcation models attempted were SVM

model with RBF and poly kernel, and Gradient Boost-

ing model (Van Messem, 2020; Stoean and Stoean,

2013; Friedman, 2006).

3 RESULTS AND DISCUSSIONS

The proposed pipeline framework (#1) extracts sEMG

features from accelerometer data which is further

supplied to classiﬁer model. The proposed pipeline

framework which transfers the modality of sensors

data is benchmarked with two staright-forward meth-

ods. One where the original sEMG features are ex-

tracted from the original sEMG signals and further

classiﬁed for upper limb movements also referred to

as Benchmark framework#1. The other one which

employs accelerometer data to classify the motions,

which is referred to as Benchmark framework#2.

These two benchmarking frameworks are illustrated

in Figures 5 (a, b). The proposed framework (#2) is

comprised of extracted synthetic sEMG features from

the accelerometer data and further merging with the

original sEMG features. The augmented dataset con-

sists of synthetic features and original features which

is then employed to train the classiﬁer model. The

overall process ﬂow is illustrated in Figure 6. All the

benchmark frameworks are compared with the pro-

posed ones. While the ﬁrst proposed framework ex-

tracts the sEMG features from a different modality

of sensors, the second proposed framework utilizes

the same but augments the dataset with original fea-

tures to build the classiﬁer model. Table 1 shows

the accuracy of the classiﬁer models including the

proposed one and the benchmark model for the four

wrist movements. The proposed synthetic features

evolved framework shows better precision than the

original features extracted proposed framework (#1)

which is expected considering the features are ex-

tracted from accelerometer data in the proposed work

which are proximally placed whereas the EMG elec-

trodes are distally placed. Additionally the proposed

framework (#2) trained with augmented synthetic and

original sEMG features shows comparable accuracy

with respect to the original ones. As expected the

accelerometer devised classiﬁer shows high precision

over others which is attributed to the proximal place-

ment of sensors, that captures the dynamic motions

accurately.

(a)

(b)

Figure 5: (a) Benchmark framework#1 where the original

features from sEMG signals are extracted and further sup-

plied to classiﬁer model, and (b) Benchmark framework#2

where the accelerometer are used to classify the motion

instead of extracting sEMG features from Accelerometer

data.

The proposed frameworks along with two other

benchmark frameworks are evaluated for Precision,

Recall, and F1-score. Table 1 lists the accuracy

metrics for the proposed model (#1), two bench-

mark frameworks and proposed merged classiﬁer

Transfer-Modal Extraction of Surface EMG Features for Upper Limb Motor Classiﬁcation

725

Figure 6: Proposed framework (#2) where the synthetic and

original sEMG features are supplied to train classiﬁer mod-

els.

Table 1: Accuracy metrics for the proposed frameworks

along with other benchmark frameworks.

Models Motions Precision (%) Recall (%)

SCORE (%)

Proposed

framework#1

Benchmark

framework#1

Benchmark

framework#2

Proposed

framework#2

model. The proposed framework (#1) which classi-

ﬁes the four wrist movements using three predicted

sEMG features derived from accelerometer readings,

achieves an average precision of 83%, with average

F1 score of 80.25% This proposed framework lever-

ages the mapping of two different modality of sensors

- accelerometer and sEMG, by predicting sEMG fea-

tures from accelerometer readings before classifying

the wrist movements. The trained prediction model

tries to map the sEMG features for an accelerom-

eter signal that is generated on wrist action. This

prediction is not real, but it forsees that sEMG fea-

tures from the accelerometer signal which is gener-

ated from movement dynamics and not from muscle

activity. The mapping of muscle activity to movement

is performed by the prediction model.

The prediction model is trained to accurately map

the accelerometer signal with sEMG features for all

the four wrist movements.

In comparison, the benchmark framework#1

shown in Figure 5 (a), classiﬁes wrist movements

through the three sEMG features directly extracted

from the original sEMG signals, achieving an average

precision of 74.5% as listed in the Table 1.

While this framework beneﬁts from the direct

measurement of muscle activity through sEMG sig-

nals, the electrodes positioned at distal end is prone

to lose critical information and hence the drop in ac-

curacy. Multiple electrodes at the distal end is bound

to improve the accuracy by consolidating vector fea-

tures.

Direct sEMG readings offer high temporal reso-

lution and are effective for muscle activity detection,

but are limited by noise and variability in electrode

placement.

The benchmark framework#2 as portrayed in Fig-

ure 5 (b), which classiﬁes wrist movements directly

from accelerometer readings, achieves the highest av-

erage precision of 94% among all the frameworks.

This framework’s great precision underscores the ef-

fectiveness of accelerometer data in capturing the dy-

namics of wrist movements. Accelerometers provide

detailed motion data, including acceleration and an-

gular velocity, which are directly related to the phys-

ical movements being classiﬁed. The high precision

indicates that accelerometer readings are highly reli-

able for distinguishing between different wrist move-

ments. However, this framework lack insights into the

underlying muscle activity, which is crucial in certain

applications such as rehabilitation or bio-mechanical

analysis. Besides, the accelerometers positioning is

always restricted to speciﬁc regions thereby limiting

the design space for rehabilitation and related tasks.

The proposed merged framework#2 as illustrated

in the Figure 6, combines the dataset of predicted

EMG features along with the dataset of original

sEMG features, thus creating a richer dataset. It is

then employed as input to classify the upper limb mo-

tions. The proposed merged dataset devised frame-

work achieves an average precision of 77.5% which

is comparable with the original sEMG feature trained

framework (74.5%). This strengthens the proposal

of augmenting the dataset through predicted sEMG

features from the other modality of sensors when-

ever the original sEMG signal and data acquisition is

not possible. It is also noted that pure synthetic fea-

tures present classiﬁer (proposed variant) accuracy of

around 83%, which is vastly deviated from the origi-

nal framework (#1). Hence the augmented dataset is

much closer to the real sEMG signal generated and

devised framework.

Additionally, the merged dataset devised model

(proposed #2) was evaluated on original unseen

sEMG features and synthetically generated but un-

seen sEMG features individually.

Accuracy of 91% is achieved when the proposed

framework 1 (which is trained completely on the orig-

inal unseen dataset) is used and is given the aug-

mented training dataset of the proposed framework

2 for testing. Accuracy of 95% is achieved when

the proposed framework 2 (which is trained on the

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

726

augmented dataset) is used and is given the origi-

nal unseen dataset to test. This strongly supports the

use of accelerometers derived sEMG features for aug-

menting the dataset and further employing the trained

model on unseen dataset successfully. The mapping

of accelerometer data to sEMG features although does

not fully capture the electro-muscular activity but it

augments the dataset well to train the model and ex-

tract discriminative features for original and unseen

sEMG features. The augmentation of sEMG dataset

through other modality of sensors opens up several

avenues to train the model for human limb. The four

wrist actions are purely to showcase the effect of syn-

thetic features and the dataset.

4 CONCLUSION

The proposed sEMG features extracted from a dif-

ferent modality of sensor is introduced in this work

towards classifying four of the upper limb motions.

These motion classiﬁers are useful to design and de-

velop an adaptive rehabilitation system with predic-

tive control strategy. The proposed synthetically gen-

erated sEMG features from the ML model show high

degree of correlation with the original sEMG gener-

ated features for four of the wrist actions. The data

collected was based on the four motions which were

restricted to the ﬂexor carpi ulnaris and extensor carpi

radialis muscles. An ensemble model was proposed

and veriﬁed for attaining maximum accuracy for ex-

traction of sEMG features. Post the prediction of the

features, RF classiﬁer was employed for classifying

the four identiﬁed motions. The machine learning

models were optimised with appropriate hyperparam-

eters and features ensuring best possible results. The

predicted sEMG features augmented with orginal fea-

tures demonstrates comparable motion classiﬁer ac-

curacy. Additionally for unseen original sEMG fea-

tures, the proposed augmented dataset trained model

showcased a superior accuracy of 91%. This is a step

for developing large dataset of human limb motion,

which will possibly give a much needed impetus for

designing a generic human limb rehabilitation system

that is applicable to the most of the needy ones. The

generic rehabilitation system will further reduce the

cost to the users. The dataset and model ﬁles are made

freely available at (Mod, ) for further use to the re-

searchers and scientiﬁc community.

REFERENCES

Dataset and model ﬁles are available at: https://drive.goog

le.com/drive/folders/1fEIuRkBEIlyJBbj79CEXPjqh

9B16Ibon?usp=sharing.

Arlot, S. and Celisse, A. (2009). A survey of cross vali-

dation procedures for model selection. Statistics Sur-

veys, 4.

Ashcroft, R. E. et al. (2008). The declaration of helsinki.

The Oxford textbook of clinical research ethics, pages

141–148.

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

Merletti, R. (2020). Surface emg in clinical assess-

ment and neurorehabilitation: Barriers limiting its use.

Frontiers in Neurology, 11.

Chandrasekhar, V., Vazhayil, V., and Rao, M. (2020a). De-

sign of a portable anthropomimetic upper limb reha-

bilitation device for patients suffering from neuromus-

cular disability. In 2020 42nd Annual International

Conference of the IEEE Engineering in Medicine &

Biology Society (EMBC), pages 4708–4712.

Chandrasekhar, V., Vazhayil, V., and Rao, M. (2020b). De-

sign of a real time portable low-cost multi-channel

surface electromyography system to aid neuromus-

cular disorder and post stroke rehabilitation patients.

In 2020 42nd Annual International Conference of the

IEEE Engineering in Medicine & Biology Society

(EMBC), pages 4138–4142.

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.

Chizari, A., Knop, T., Sirmacek, B., van der Heijden, F., and

Steenbergen, W. (2020). Exploration of movement

artefacts in handheld laser speckle contrast perfusion

imaging. Biomedical optics express, 11(5):2352–

2365.

Eschweiler, J., Li, J., Quack, V., Rath, B., Baroncini, A.,

Hildebrand, F., and Migliorini, F. (2022). Anatomy,

biomechanics, and loads of the wrist joint. Life, 12(2).

Farag, W. (2020). Kalman-ﬁlter-based sensor fusion applied

to road-objects detection and tracking for autonomous

vehicles. Proceedings of the Institution of Mechani-

cal Engineers, Part I: Journal of Systems and Control

Engineering, 235:095965182097552.

Friedman, J. H. (2006). Recent advances in predictive (ma-

chine) learning. Journal of classiﬁcation, 23(2):175–

197.

Frost, J. (2019). How to interpret r-squared in regression

analysis. statistics by jim, making statistics intuitive.

Statistics by Jim Making Statistics Intuitive.

Hiengkaew, V., Jitaree, K., and Chaiyawat, P. (2012). Mini-

mal detectable changes of the berg balance scale, fugl-

meyer assessment scale, timed “up & go” test, gait

speeds, and 2-minute walk test in individuals with

chronic stroke with different degrees of ankle plan-

tarﬂexor tone. Archives of Physical Medicine and Re-

habilitation, 93(7):1201–1208.

InvenSense, I. (2013). Mpu-6000 and mpu-6050 product

speciﬁcation revision 3.4. United States.

Transfer-Modal Extraction of Surface EMG Features for Upper Limb Motor Classiﬁcation

727

Ioffe, S. and Szegedy, C. (2015). Batch normalization: Ac-

celerating deep network training by reducing internal

covariate shift. In International conference on ma-

chine learning, pages 448–456. pmlr.

Ismail Fawaz, H., Forestier, G., Weber, J., Idoumghar, L.,

and Muller, P.-A. (2018). Data augmentation using

synthetic data for time series classiﬁcation with deep

residual networks.

Jonna, P., Madurwar, A., BK, A., and Rao, M. (2024). De-

sign of 6-dof holonomic drive-based upper and lower-

limb stroke rehabilitation system. In 2024 10th In-

ternational Conference on Automation, Robotics and

Applications (ICARA), pages 233–239.

Jonna, P. and Rao, M. (2022). Design of a 6-dof cost-

effective differential-drive based robotic system for

upper-limb stroke rehabilitation. In 2022 44th Annual

International Conference of the IEEE Engineering in

Medicine & Biology Society (EMBC), pages 1423–

1427.

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W.,

Ye, Q., and Liu, T.-Y. (2017). Lightgbm: A highly

efﬁcient gradient boosting decision tree. Advances in

neural information processing systems, 30.

Khalilia, M., Chakraborty, S., and Popescu, M. (2011). Pre-

dicting disease risk from highly imbalanced data using

random forest. BMC medical informatics and decision

making, 11:51.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012).

Imagenet classiﬁcation with deep convolutional neu-

ral networks. In Pereira, F., Burges, C., Bottou, L.,

and Weinberger, K., editors, Advances in Neural In-

formation Processing Systems, volume 25. Curran As-

sociates, Inc.

Kwakkel, G., Kollen, B. J., and Krebs, H. I. (2008). Effects

of robot-assisted therapy on upper limb recovery after

stroke: A systematic review. Neurorehabilitation and

Neural Repair, 22(2):111–121. PMID: 17876068.

Lawrence, E. L., Fassola, I., Werner, I., Leclercq, C., and

Valero-Cuevas, F. J. (2014). Quantiﬁcation of dexter-

ity as the dynamical regulation of instabilities: Com-

parisons across gender, age, and disease. Frontiers in

Neurology, 5.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-

ing. nature, 521(7553):436–444.

Loureiro, R. C. and Harwin, W. S. (2007). Reach & grasp

therapy: Design and control of a 9-dof robotic neuro-

rehabilitation system. In 2007 IEEE 10th Interna-

tional Conference on Rehabilitation Robotics, pages

757–763.

Mason, S. and Birch, G. (2003). A general framework for

brain-computer interface design. IEEE Transactions

on Neural Systems and Rehabilitation Engineering,

11(1):70–85.

Merletti, R. and Cerone, G. L. (2020). Tutorial. surface emg

detection, conditioning and pre-processing: Best prac-

tices. Journal of Electromyography and Kinesiology,

54:102440.

Novak, D., Mihelj, M., and Munih, M. (2012). A survey of

methods for data fusion and system adaptation using

autonomic nervous system responses in physiological

computing. Interacting with Computers, 24(3):154–

172.

P, J., GVK, S., Rao, M., Bapat, J., and Das, D. (2023). Xore-

hab: Iot enabled wheelchair based lower limb reha-

bilitation system. In 2023 45th Annual International

Conference of the IEEE Engineering in Medicine &

Biology Society (EMBC), pages 1–5.

Pal, M. (2005). Random forest classiﬁer for remote sensing

classiﬁcation. International journal of remote sensing,

26(1):217–222.

Phinyomark, A., Nuidod, A., Phukpattaranont, P., and Lim-

sakul, C. (2012). Feature extraction and reduction of

wavelet transform coefﬁcients for emg pattern classiﬁ-

cation. Elektronika ir Elektrotechnika, 122(6):27–32.

Reddy, M., Jonna, P., Perala, S., Rao, M., and Vazhiyal, V.

(2023). Automated microsurgical tool categorization

using a surface-based emg system. In 2023 45th An-

nual International Conference of the IEEE Engineer-

ing in Medicine & Biology Society (EMBC), pages 1–

Schreuders, T., Brandsma, J., and Stam, H. (2019). Func-

tional Anatomy and Biomechanics of the Hand, pages

3–21.

Shi, Y. (2012). The application of the butterworth low-

pass digital ﬁlter on experimental data processing. In

2011 International Conference in Electrics, Commu-

nication and Automatic Control Proceedings, pages

225–230. Springer.

Stoean, R. and Stoean, C. (2013). Modeling medical deci-

sion making by support vector machines, explaining

by rules of evolutionary algorithms with feature selec-

tion. Expert Systems with Applications, pages 2677–

2686.

Upside Down Labs (2024). Muscle-bioamp-patchy. https:

//github.com/upsidedownlabs/Muscle-BioAmpPat

chy/tree/main/hardware. [Online; accessed 11-June-

2024].

Van Messem, A. (2020). Chapter 10 - support vector ma-

chines: A robust prediction method with applications

in bioinformatics. In Srinivasa Rao, A. S. and Rao,

C., editors, Principles and Methods for Data Science,

Handbook of Statistics, pages 391–466.

Vinay, K., Nagaraj, K., Arvinda, H. R., Vikas, V., and Rao,

M. (2021). Design of a device for lower limb pro-

phylaxis and exercise. IEEE Journal of Translational

Engineering in Health and Medicine, 9:1–7.

Vinay, K., Vazhayil, V., and Rao, M. (2022). An event

driven approximate bio-electrical model generating

surface electromyography rms features. In 2022 35th

International Conference on VLSI Design and 2022

21st International Conference on Embedded Systems

(VLSID), pages 204–209.

Vitali, R. V. and Perkins, N. C. (2020). Determin-

ing anatomical frames via inertial motion capture:

A survey of methods. Journal of Biomechanics,

106:109832.

Zhou, Z.-H. (2012). Ensemble methods: foundations and

algorithms. CRC press.

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

728