Dimensionality Reduction and Classification Methods

for High-Accuracy EMG Signal Interpretation in

Prosthetics

Chiang Liang Kok Chee Kit Ho Yit Yan Koh

College of Engineering Engineering Cluster College of Engineering
The University of Newcastle Singapore Institute of Technology The University of Newcastle
Australia Singapore Australia
[email protected] [email protected] [email protected]

Fu Kai Tan Tee Hui Teo

School of Science and Technology Engineering Product Development
Singapore University of Social Science Singapore University of Technology
[email protected] and Design
[email protected]
TENCON 2024 - 2024 IEEE Region 10 Conference (TENCON) | 979-8-3503-5082-1/24/$31.00 ©2024 IEEE | DOI: 10.1109/TENCON61640.2024.10902895

Abstract— Signals are crucial in conveying information across unique EMG signals from 43 participants performing 16
various fields. This paper presents new methods for gestures, captured seven times each. Using signal
processing electromyographic (EMG) signals to create AI processing techniques like Mean Absolute Value (MAV),
systems that decode muscle activity for arm movement Waveform Length (WL), Zero Crossing (ZC), Slope Sign
control. Using an advanced dataset, the study focuses on Change (SSC), Root Mean Square (RMS), Energy of
enhancing prosthetic control and rehabilitation technologies Wavelet Coefficients (EWLs), and Enhanced Mean
through sophisticated signal processing and machine learning Absolute Value (EMAV), the study aims to uncover
techniques. Various preprocessing steps improved signal patterns in EMG signals indicating forearm muscle
quality, and a diverse set of features was extracted and
activity. Dimensionality reduction techniques like Linear
classified. The results highlight the potential for more
intuitive and responsive robotic arm movements, contributing
Discriminant Analysis (LDA) and Principal Component
to better prosthetic and rehabilitation solutions. Analysis (PCA) will refine features, enhancing their
discriminative power. Machine learning models will
Keywords—machine learning; signal processing;
predict forearm muscle movements, advancing intuitive
electromyography (EMG) prosthetic control systems and rehabilitation technologies.
The research involves preprocessing raw EMG signals,
I. INTRODUCTION decomposing and reconstructing them using the
Biorthogonal 3.3 Discrete Wavelet Transform (DWT), and
Signals are fundamental to science, technology, and
implementing feature extraction and dimensionality
communication, acting as carriers of information through
reduction. Classification algorithms will predict hand
changing patterns, amplitudes, and frequencies. In deep
gestures from the optimized dataset, focusing on five
learning, accurately modeling these signals is crucial for
primary gestures: Little Finger Extension (LFE), Index
making predictions, requiring identification of key
Finger Extension (IFE), Thumb Extension (TE), Hand
algorithm factors and understanding relationships between
Open (HO), and Hand Close (HC). Despite potential
signals and features. Surface Electromyography (SEMG)
limitations, such as dataset quality and EMG signal
for Hand Gesture Recognition (HGR) faces limitations due
variability, this research aims to enhance understanding of
to individual EMG signal variability, necessitating
forearm muscle signals and improve gesture prediction
personalized HGR systems [1]. Combining SEMG with
accuracy, contributing to assistive technology
electroencephalography (EEG) signals, using techniques
advancement.
like Graph Convolutional Networks (GCNs) and
Functional Connectivity enhances arm motion recognition II. LITERATURE REVIEW
accuracy [2]. Advances in technology enable the non-
invasive capture of high-definition SEMG signals, integral The literature on machine learning-based feature
to Human–Machine Interfaces (HMI) and biocybernetics. extraction and classification of EMG signals for intuitive
For individuals without hands, EMG technology enables prosthetic control has expanded significantly in recent
prosthetic control via muscle activity detection, translating years, reflecting the growing interest in developing
muscle signals into commands for prosthetic hands, advanced assistive technologies. Traditional approaches to
enhancing autonomy and quality of life. This research EMG signal analysis often relied on basic statistical
utilizes advanced signal feature extraction and machine methods and manual feature selection, which posed
learning to analyze the GRABMyo dataset, containing limitations in capturing the complex patterns inherent in
EMG signals. Recent advancements have introduced more

9979-8-3503-5082-1/24/$31.00 ©2024 IEEE 1845

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on April 05,2025 at 12:45:25 UTC from IEEE Xplore. Restrictions apply.
sophisticated techniques, such as time–frequency analysis III. OVERVIEW OF SIGNAL PROCESSING METHODS
and wavelet transform, to better characterize EMG signals. Signal processing [12-14] is essential for analyzing and
For instance, this work [3] emphasized the importance of interpreting raw data by extracting meaningful
feature extraction methods like Mean Absolute Value information, reducing noise, and identifying patterns from
(MAV), Zero Crossing (ZC), and Root Mean Square input signals. This process enhances data quality, enables
(RMS) for enhancing signal interpretation. These methods, efficient storage and transmission, and facilitates
when combined with machine learning algorithms, have subsequent analysis and interpretation in various
shown promising results in classifying various hand applications.
gestures and muscle movements. Machine learning [3-4]
particularly deep learning, has revolutionized EMG signal A. Average Referenciing
processing by automating feature extraction and improving Average referencing, commonly used in EEG and EMG
classification accuracy. Convolutional Neural Networks signal processing, enhances signal accuracy by removing
(CNNs) and Recurrent Neural Networks (RNNs) have been common noise across channels. This technique calculates
employed to learn spatial and temporal dependencies in the average of all channels at each time point and subtracts
EMG data, leading to significant improvements in gesture it from each individual channel, improving the signal-to-
recognition tasks. This work [4] demonstrated that CNNs noise ratio and data quality. In the GRABMyo dataset, with
could effectively capture the spatial patterns of EMG 28 electrodes capturing forearm signals, average
signals, resulting in a higher classification performance referencing is highly effective. Applied to the 16 forearm
compared to traditional methods. Similarly, the paper [5] and 12 wrist channels, it reduces common noise and aligns
utilized Support Vector Machines (SVMs) and Random signals around a zero-reference point, ensuring
Forests to classify EMG signals from the Ninapro database, consistency. This is crucial for analyzing interactions
achieving high accuracy and robustness. The integration of among the 28 electrodes and mitigates uncorrelated or
EMG with other bio signals, such as random noise, improving signal clarity and accuracy. The
electroencephalography (EEG), has also been explored to frequency range of EMG signals varies with the type of
enhance prosthetic control systems. Other research [6–9] gesture, as different gestures activate various muscle
investigated the fusion of SEMG and EEG signals using groups with distinct frequencies. Generally, EMG signals
Graph Convolutional Networks (GCNs) to improve the range from 10 Hz to 450 Hz, with higher frequencies
recognition of complex arm movements. This multi-modal indicating more forceful or rapid muscle contractions. In
approach leverages complementary information provided some cases, or pathological conditions, frequencies can
by different types of bio signals, leading to more reliable exceed this range. In biomedical signal recording, a notch
and flexible control mechanisms. Additionally, feature filter is used to eliminate power line noise, typically
selection and dimensionality reduction techniques, such as between 50 Hz and 60 Hz. This noise can interfere with
Principal Component Analysis (PCA) and Linear signals like EMG, which also fall within this frequency
Discriminant Analysis (LDA), have been applied to range. The notch filter targets and removes power line
optimize the features extracted from EMG signals, noise without affecting the muscle activity in the EMG
reducing computational complexity while maintaining high signal. During the capture of EMG signals, DC components
classification accuracy. The application of these advanced may arise due to factors like electrode offset or baseline
techniques has significant implications for the drift. These DC components, typically at 0 Hz, can interfere
development of intuitive prosthetic control systems. By with analyzing changes in muscle activity over time. To
enabling a more accurate and responsive control of address this, a high-pass filter with a cutoff frequency of
prosthetic limbs, these methods can substantially improve 0.1 Hz is applied. This filter attenuates or removes low-
the quality of life for individuals with limb loss. Future frequency DC components while preserving higher-
research is likely to focus on further refining machine frequency components that represent muscle activity. This
learning models [6-9], exploring new feature extraction helps eliminate DC offset or baseline variations that could
methods, and integrating additional bio signals to enhance distort information about muscle contractions, ensuring
the functionality and usability of prosthetic devices. accurate readings. Figure 2 depicts the signals after
Overall, the convergence of machine learning and EMG undergoing filtering, preparing them for analysis using the
signal processing represents a promising frontier in the Discrete Wavelet Transform (DWT). Filtering is essential
field of assistive technology, with the potential to deliver for improving signal quality by removing noise, artifacts,
more natural and effective prosthetic control solutions [12- and extraneous components, ensuring the DWT application
13]. Figure 1 depicts the comprehensive process diagram is effective and precise.
outlining the distinct phases of the research from Signal
Processing to Machine Learning.

Figure 2. PSD analysis post filtering of EMG signals captured from the
Figure 1. Overall process diagram for the approach of the research. forearm.

1846 TENCON 2024 – 2024 IEEE Region 10 Conference (TENCON)

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on April 05,2025 at 12:45:25 UTC from IEEE Xplore. Restrictions apply.
B. Discrete Wavelet Transform—Biorthogonal 3.3 wavelet-transformed signals, supporting tasks like signal
The Fourier transform, while powerful, is inefficient for classification and denoising. Enhanced Mean Absolute
representing abrupt changes because it uses infinite Value (EMAV) improves traditional MAV by
sinusoids without considering time or spatial localization. incorporating weighting factors to emphasize higher
Time–frequency domain representation is more effective amplitude signal components, enhancing accuracy in
for non-stationary signals. Wavelet analysis offers two representing muscle activity in EMG signal processing.
main transforms: discrete wavelet transform (DWT) and Linear Discriminant Analysis (LDA) is a statistical method
continuous wavelet transform (CWT), which differ in used for dimensionality reduction and classification tasks.
wavelet scaling and shifting. DWT is favored for It seeks to find a linear combination of features that
bioelectrical signal analysis due to its ability to decompose effectively separates multiple classes in a dataset by
signals into different frequency bands, preserving crucial minimizing within-class variability and maximizing
information. DWT's dual filter bank structure balances between-class variability.
time and frequency localization, enhancing feature IV. RESULTS AND DISCUSSION
extraction and denoising. This research investigates
DWT’s applications in biomedical signal processing and A. Performance Metrics for GrabMYO dataset
image compression. The DWT process involves filtering Figure 4 presents both the scatter plot, which visually
the signal with low-pass and high-pass filters, followed by compares the GRABMyo dataset after dimensional
down sampling, resulting in high-pass (detail coefficients) reduction, highlighting clustering differences, and the
and low-pass (approximate coefficients) components. The confusion matrix evaluating the SVM model's predictions
iterative filtering process with multiple filter banks allows across different classes. The matrix shows correct
for multiple levels of decomposition, splitting the signal predictions on the diagonal and misclassifications off-
into approximate (low-frequency) and detail (high- diagonally, offering a comprehensive view of model
frequency) components. performance on the full GRABMyo dataset.

Figure 4. Scatter plot and confusion matrix illustrating SVM model

performance from the GRABMyo dataset.
Figure 3. Discrete Wavelet Transform decomposition, analysis form, with
4 levels of decomposition. Table 1. Performance metrics of the LDA for machine learning models,
KNN, NB, and the SVM, using the GRABMyo dataset
C. Signal Feature Extraction Methods Evaluation Macro
Classes (Five Hand Gestures)
According to the cited journal findings, feature Metrics Average

extraction techniques that yield high accuracy include Support Vector Machine

Mean Absolute Value (MAV), Waveform Length (WL), True Positive 94 100 98 105 98 99
Zero Crossing (ZC), Slope Sign Change (SSC), Root Mean False Positive 20 32 42 17 39 30
Square (RMS), Energy of Wavelet Coefficients (EWLs), False Negative 35 29 31 24 31 30
and Enhanced Mean Absolute Value (EMAV). MAV
True Negative 496 484 474 499 477 486
calculates the average magnitude of a signal by averaging
the absolute values of its data points, commonly used in Precision 0.82456 0.75758 0.7 0.86066 0.71533 0.77162

EMG analysis to assess muscle activity. Waveform Length Sensitivity 0.72868 0.77519 0.75969 0.81395 0.75969 0.76744
(WL) quantifies the cumulative length of waveform Accuracy 0.91473 0.90543 0.88682 0.93643 0.89147 0.90698
oscillations by summing distances between successive F-measure 0.77366 0.76628 0.72862 0.83665 0.73684 0.76841
peaks or troughs in the EMG signal. Zero Crossing (ZC)
identifies points in a signal where the amplitude changes its
sign, indicating crossings of the zero-amplitude axis. Slope Figure 5 presents the results from the LDA–SVM
Sign Change (SSC) counts changes in signal slope within a model across the GRABMyo dataset, showing accuracy
specified interval, aiding in detecting transitions in signal improved by 3.72% to 90.69% with a sample size increased
dynamics. Root Mean Square (RMS) calculates the average from 215 to 645. The GRABMyo dataset offers distinct
power of a signal by taking the square root of the average advantages over other surface EMG datasets, notably in its
of the squares of all signal values, providing insights into diversity of gestures and participant backgrounds. This
signal intensity and useful in noise analysis and power variety supports the development of robust models that
estimation. Energy of Wavelet Coefficients (EWLs) generalize well across different movements. Additionally,
measures energy distribution across frequency bands in the dataset's varied recording conditions, encompassing

TENCON 2024 – 2024 IEEE Region 10 Conference (TENCON)

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on April 05,2025 at 12:45:25 UTC from IEEE Xplore. Restrictions apply.
1847
diverse muscle contraction levels and environments, prostheses: State of the art and challenges for clinical
contribute to resilient model development. use. J. Rehabil. Res. Dev. 2011, 48, 643–659.
[9] Palaniappan, K., Kok, C.L., Kato, K. (2021). Artificial
Intelligence (AI) Coupled with the Internet of Things
(IoT) for the Enhancement of Occupational Health and
Safety in the Construction Industry. In: Ahram, T.Z.,
Karwowski, W., Kalra, J. (eds) Advances in Artificial
Intelligence, Software and Systems Engineering.
AHFE 2021. Lecture Notes in Networks and Systems,
Figure 5. Comparing performance metrics for SVM classifiers utilizing vol 271. Springer, Cham. https://doi.org/10.1007/978-
data from one session (215 samples) and the complete dataset of three 3-030-80624-8_4
sessions (645 samples) of data. [10] Nagel, S. Towards a Home-Use BCI: Fast
Asynchronous Control and Robust Non-Control
V. CONCLUSION State Detection. ResearchGate, December 2019.
The LDA–SVM model demonstrated superior accuracy Available online:
in classifying EMG signals for prosthetic control. Starting https://www.researchgate.net/publication/33842358
with 86.98% accuracy from 215 samples, it improved to 5_Towards_a_home-
90.69% when using 645 samples, highlighting the benefits use_BCI_fast_asynchronous_control_and_robust_n
of a larger dataset. This improvement underscores the on-control_state_detection (accessed on)
model’s robustness and ability to generalize across [11] Strategist, C. Gesture Control Your Life with the
different sessions and participants in the GRABMyo Myo Armband. Bahrain This Week, 8 September 2017.
dataset. Metrics such as precision, sensitivity, and F- Available online:
measure indicated high performance across various hand
https://www.bahrainthisweek.com/gesture-control-
gestures. SVM consistently outperformed KNN [15-16]
your-life-with-the-myo-armband/ (accessed on).
and NB models, especially in managing high-dimensional
data and enhancing class separation. [12] Kanade, V. What Is Machine Learning?
Understanding Types & Applications. Spiceworks, 30
REFERENCES August 2022. Available online:
[1] Pradhan, A.; He, J.; Jiang, N. Multi-day dataset of https://www.spiceworks.com/tech/artificial-
forearm and wrist electromyogram for hand gesture intelligence/articles/what-is-ml/ (accessed on).
recognition and biometrics. Sci. Data 2022, 9, 1–10. [13] Pelley, S. Feel again: Advancements in Prosthetics
https://doi.org/10.1038/s41597-022-01836-y Limb Technology Allow Feeling, Control. CBS News,
[2] Yang, S.; Li, M.; Wang, J. Fusing sEMG and EEG to 16 July 2023. Available online:
Increase the Robustness of Hand Motion Recognition https://www.cbsnews.com/news/advancements-in-
Using Functional Connectivity and GCN. IEEE Sens. J. prosthetics-limb-technology-allow-feeling-control-
2022, 22, 24309–24319. 60-minutes-transcript-2023-07-16/ (accessed on).
https://ieeexplore.ieee.org/document/9953952/ [14] Kok, C.L.; Ho, C.K.; Tan, F.K.; Koh, Y.Y. Machine
[3] Phinyomark, A.; Quaine, F.; Charbonnier, S.; Serviere, Learning-Based Feature Extraction and Classification
C.; Tarpin-Bernard, F.; Laurillau, Y. Feature extraction of EMG Signals for Intuitive Prosthetic Control. Appl.
of the first difference of EMG time series for EMG Sci. 2024, 14, 5784. doi: 10.3390/app14135784
[15] Xiang, M.; Teo, T. H.; A Practical Approach to
pattern recognition. Comput. Methods Programs Biomed.
2014, 117, 247–256. Mitigate the Dependencies of Generative AI in
[4] Park, E.J.; Lee, S.H. EMG pattern recognition based on Engineering Education (May 21, 2024). Available at
artificial intelligence techniques. IEEE Trans. Ind. SSRN: https://ssrn.com/abstract=4855880 or
Electron. 2017, 62, 778–784. http://dx.doi.org/10.2139/ssrn.4855880
[5] Atzori, M.; Gijsberts, A.; Kuzborskij, I.; Elsig, S.; Hager, [16] Teo, T. H.; Xiang, M.; Zhai, X; Ouyang, F.; ‘Using
A.-G.M.; Deriaz, O.; Castellini, C.; Muller, H.; Caputo, Artificial Intelligence Tools in product design course’,
B. Characterization of a Benchmark Database for The Asian Conference on Education (ACE2023).
Myoelectric Movement Classification. IEEE Trans.
Neural Syst. Rehabilitation Eng. 2014, 23, 73–83.
[6] Xu, R.; et al. Enhanced hand motion recognition using
effective fusion of surface electromyography and
ultrasonography imaging. IEEE Trans. Biomed. Eng.
2017, 64, 1927–1935.
[7] He, J.; et al. Dimensionality reduction of sEMG signals
using PCA and LDA. J. Med. Syst. 2015, 39, 35.
[8] Scheme, E.; Englehart, K. Electromyogram pattern
recognition for control of powered upper-limb

1848 TENCON 2024 – 2024 IEEE Region 10 Conference (TENCON)

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on April 05,2025 at 12:45:25 UTC from IEEE Xplore. Restrictions apply.