Upper Limb Movement Recognition Utilising EEG and EMG Signals for Rehabilitative Robotics

Zihao Wang, Ravi Suppiah, Suppiah, RaviLecture Notes in Networks and Systems 2023 · DOI · OpenAlex

Abstract

Upper limb movement classification is a critical component in controlling rehabilitative robotics, mapping biological signals to target activities. While Electromyography (EMG) and Electroencephalography (EEG) are widely used, their effectiveness is often hampered by environmental noise and patient-specific physiological limitations. This paper explores various feature extraction and machine learning techniques, proposing a novel decision-level multisensor fusion approach to integrate EEG and EMG signals. By testing on the WAY-EEG-GAL dataset, the study demonstrates a robust system capable of accurately predicting user intent even under suboptimal signal conditions.

Introduction

Robotic prosthetics and exoskeletons rely on interpreting a user's intent to move. Traditionally, this is achieved through Electromyography (EMG), which measures electrical activity from muscles. However, EMG has significant drawbacks: it can be extremely weak in patients with severe muscle damage or neuromuscular disorders and is highly susceptible to real-world electrical noise.

Electroencephalography (EEG), recording brain waves directly, offers a solution by tapping into the source of intent. Yet, EEG signals are notoriously "messy" and difficult to classify accurately on their own. This research addresses the need for a robust, multimodal system that integrates both signal types to provide a reliable control mechanism for upper-limb rehabilitation.

Methodology

The proposed system utilizes a comprehensive pipeline involving feature engineering, deep learning, and a unique fusion algorithm:

  1. Feature Extraction:
    • EEG: Seven key channels from the central scalp region were selected (see Figure 2). Extracted features included time-domain (standard deviation), frequency-domain (power spectral density), and time-frequency features (Discrete Wavelet Transform).
  2. EMG: Focusing on five muscle groups, five time-domain features were extracted, including Mean Absolute Value and Waveform Length.
  3. Feature Selection: The Boruta algorithm (a random forest-based method) was employed to prune irrelevant data, reducing 70 potential EEG features to the 51 most discriminative ones.
  4. Classification: Several models were compared, including Support Vector Machines (SVM) and Multi-layer Perceptrons (MLP). Long Short-Term Memory (LSTM) networks proved most effective for handling the temporal nature of these biological signals.
  5. Decision-Level Fusion: The core innovation is a fusion algorithm that calculates a "Weightage" based on the expected accuracy of each sensor and a "Degree of Noisiness" based on real-time signal fluctuations. The final prediction prioritizes the cleaner, more reliable sensor at any given moment.
Standard 32-channel EEG electrode placement map used for selecting informative brain regions.

Experimental Setup

The study utilized the WAY-EEG-GAL dataset, containing simultaneous EEG and EMG recordings from 12 participants performing a "grasp-and-lift" task. Participants reached for an object, lifted it (with varying weights and surface frictions), held it, and returned it. This allowed for categorization into four distinct stages: resting, extension, lifting, and flexion.

Discussion and Key Insights

  • Multimodal Robustness: While EMG is generally more accurate, it fails under noise. Integrating EEG provides a secondary source of intent that can maintain control when muscle signals are compromised.
  • LSTM Performance: For both signal types, LSTM networks significantly outperformed traditional classifiers like k-Nearest Neighbors (kNN) or SVMs, highlighting the importance of temporal context.
  • EEG Complexity: Visualizing EEG data in the time domain (Figure 3) reveals no obvious patterns, necessitating complex frequency-domain analysis to extract usable features.
  • Channel Requirements: Reducing the EEG setup to a single channel (C3) caused accuracy to drop from 82.7% to 48.5%, suggesting that a minimum of seven channels is necessary for reliable recognition.
Time-domain visualization of raw EEG signals from the C3 channel, demonstrating high noise levels.

Results

  • Baseline Accuracy: On clean data, the LSTM model achieved 82.7% accuracy for EEG and 99.8% for EMG.
  • Noise Tolerance: With added Gaussian noise, standalone EEG accuracy dropped to 65.3%. However, the fusion algorithm maintained 89.7% accuracy by intelligently shifting weight to the less-contaminated signal.

Conclusion

The results demonstrate that decision-level fusion is a viable path for creating rehabilitative robots that work for a wider range of patients. Future work will address individual biological variations (e.g., muscle fatigue) and explore adding more sensors like Electrocardiography (ECG) to further increase reliability in complex environments.

Summary updated

BibTeX
@inproceedings{wang2023upper,
  author = {Zihao Wang and Ravi Suppiah and Suppiah, Ravi},
  title = {{Upper Limb Movement Recognition Utilising EEG and EMG Signals for Rehabilitative Robotics}},
  booktitle = {Lecture Notes in Networks and Systems},
  year = {2023},
  doi = {10.1007/978-3-031-28076-4\_49},
}
Figure