Electromyogram-Based Classification of Hand and Finger Gestures Using Artificial Neural Networks

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Electromyogram (EMG) signals are increasingly utilized for hand and finger gesture recognition.
• Existing research primarily focuses on wrist and whole-hand gestures, overlooking the complexity of individual finger (IF) gestures.

Purpose of the Study:

• To develop and evaluate EMG-based classifiers for recognizing individual finger gestures using machine learning.
• To assess the performance of different machine learning algorithms (ANN, SVM, RF, LR) for IF gesture classification.
• To investigate the effectiveness of time-domain (TD) features with fixed electrode placement.

Main Methods:

• Ten healthy subjects performed ten gestures, including seven IF gestures.
• EMG signals were recorded from three channels, and six TD features were extracted per channel.
• Personalized classifiers were built using Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR).

Main Results:

• The ANN achieved the highest mean accuracy (0.940), followed by SVM (0.876), RF (0.831), and LR (0.539).
• ANN demonstrated the lowest inter-subject variability in accuracy, indicating robustness against individual differences.
• The proposed method achieved a favorable gesture-to-channel ratio using only TD features.

Conclusions:

• The ANN classifier shows superior performance for EMG-based individual finger gesture recognition.
• The developed method offers a promising approach for enhancing system usability and reducing computational load in gesture recognition.
• Fixed electrode placement and TD features provide an effective strategy for challenging IF gesture classification.