sEMG-Based Gesture Recognition Using Temporal History.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Surface electromyography (sEMG) pattern recognition faces challenges with overlapping classes, limiting classification accuracy.
• Current learning-based methods struggle to separate these overlapping sEMG patterns using nonlinear decision boundaries.

Purpose of the Study:

• To introduce a novel post-processing method for enhancing sEMG pattern recognition by leveraging temporal history.
• To improve classification accuracy by adjusting errors in overlapping sEMG patterns.

Main Methods:

• A post-processing technique was developed to analyze the temporal history of sEMG patterns.
• The method calculates instantaneous pattern separability to confirm prediction confidence (reliability).
• Unreliable predictions are adjusted using previously stored reliable results, integrated as a back-end step for existing classifiers (MLC, KNN, SVM).

Main Results:

• Overall classification accuracy improved by 8.12%p (MLC), 7.68%p (KNN), and 11.63%p (SVM).
• Accuracy enhancement before gesture completion reached 4.23%p (MLC), 4.23%p (KNN), and 11.12%p (SVM).
• The method demonstrated significant accuracy gains across different classifiers for 13 hand gestures, including overlapping patterns.

Conclusions:

• The proposed post-processing method effectively enhances sEMG pattern recognition by addressing overlapping classes through temporal analysis.
• This approach leads to more accurate and faster hand gesture classification, even before gesture completion.
• The technique is versatile, applicable to various classifiers without altering their core algorithms.