ML/DL Techniques to analyze EMG signal across various domains: Exhaustive Review

Abstract

Research on Machine Learning (ML) and Deep Learning (DL) applied to Electromyography (EMG) signal analysis
shows widespread use in medical and rehabilitation applications for prosthetic control, diagnosis of neuromuscular
disorders, and gait analysis, as well as in Human-Computer Interaction (HCI) for gesture recognition and user input
systems.
Purpose
The central goal of employing ML and DL for EMG analysis is to accurately and robustly decode human motor intent
and neuromuscular activity. Applications include controlling prosthetic devices, diagnosing neuromuscular disorders,
guiding rehabilitation, and enabling human-computer interaction.
Methods
The analysis aims to overcome inherent challenges of EMG signals, such as noise, variability due to electrode
placement, and changes with muscle fatigue, to achieve reliable, real-time performance.
The standard methodology involves a multi-stage pipeline:
1. Data Acquisition: Non-invasive surface EMG (sEMG) using electrodes placed on the skin over target
muscles. Invasive intramuscular EMG (iEMG) offers higher selectivity for non-clinical applications.
2. Preprocessing: Raw EMG signals are noisy and require filtering to remove artifacts like power-line
interference, motion artifacts, and baseline drift, band-pass filtering (e.g., 10–500 Hz), notch filtering (e.g.,
at 50/60 Hz), rectification, and segmentation.
3. Feature Extraction: from frequency domain (e.g., Mean/Median Frequency), or time-frequency domain
(e.g., Wavelet Transform coefficients), the Deep learning models, in contrast, and extract these high-level
features automatically from raw or minimally processed data.
4. Classification/Regression: using ML/DL models for either discrete classification (e.g., gesture
recognition) or continuous regression (e.g., joint angle or force estimation).
Common techniques include Random Forests, Convolution Neural Networks (CNNs), and Recurrent
Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and Support Vector Machines
(SVM)
5. Evaluation: Models are evaluated using metrics with ongoing research focused on improving accuracy,
reducing latency, handling data imbalance, and enabling real-time control on embedded systems.
Results
• High Accuracy: Both ML and DL models with very high accuracy in controlled settings, with some
applications classification accuracies exceeding 99%.
• DL Advances: Deep learning models, particularly hybrid CNN-LSTM architectures automatically
learning more abstract and robust features more efficient.
• Real-time Performance: DL models have been optimized for real-time applications in fields like
prosthetic control, enabling responsive and intuitive human-machine interfaces.
• Domain-Specific Outcomes:
o Prosthetics and Robotics: DL classifies hand and arm gestures, allowing more naturalistic
and precise control of robotic hands and exoskeletons.
o Clinical Diagnostics: healthy muscle activity and signals from patients with neuromuscular
disorders like myopathy and Amyotrophic Lateral Sclerosis (ALS), achieving high precision.
o Rehabilitation: provides biofeedback for guided rehabilitation exercises.
These advanced methods, combined with techniques like feature extraction in the time and frequency domains (e.g.,
RMS, MAV, Wavelet Transform) and noise reduction, allow for accurate, real-time detection and classification of
muscle activity, leading to more intuitive and responsive control systems.
Conclusion
ML and DL techniques have revolutionized EMG signal analysis, enabling accurate and robust decoding of human
intent across diverse domains especially in complex, real-world scenarios. The continued development of these
techniques, alongside improved data acquisition hardware, promises more intuitive and effective human-machine
interfaces for prosthetics, advanced clinical diagnostics, and targeted rehabilitation therapies.