An Exploratory Study of Single Channel Surface Electromyography for Hand Gesture Classification

Technical Summary: An Exploratory Study of Single Channel Surface Electromyography for Hand Gesture Classification

Problem Statement
Accurate hand gesture recognition using surface electromyography (sEMG) typically relies on dense, multi-channel sensor arrays and computationally intensive machine learning models. While effective, these systems present significant barriers to practical deployment in low-power, embedded, and consumer-grade applications due to high hardware complexity, power consumption, and cost. Furthermore, models trained on high-dimensional, multi-channel data often require substantial computational resources, making real-time inference on edge devices challenging. This study addresses the feasibility of achieving accurate hand gesture classification using only a single sEMG channel combined with lightweight machine learning architectures, aiming to facilitate cost-effective, low-power gesture-recognition systems.

Methodology
The research utilized a dataset originally collected by Khushaba et al., comprising sEMG signals from eight healthy subjects performing ten distinct hand gestures (five individual finger flexions and five combined pinching/closing gestures). The study focused exclusively on Channel 1, acquired from the volar (flexor) muscles of the right forearm, as flexor signals are known to be more discriminative for hand gestures.

The experimental pipeline involved the following stages:

1. Signal Preprocessing: Raw signals (sampled at 4000 Hz) were bandpass filtered (20–450 Hz) and notch-filtered to remove 50 Hz interference. Data was segmented into overlapping windows of 300 ms with a 50 ms overlap.
2. Feature Extraction: A comprehensive set of 32 features was extracted per window, spanning multiple domains:
   • Time-domain: Integrated EMG (IEMG), Mean Absolute Value (MAV), Root Mean Square (RMS), Variance (VAR), Waveform Length (WL), Zero Crossings (ZC), Simple Square Integral (SSI), Log Detector (LOG), Willison Amplitude (WAMP), Maximum Fractal Length (MFL), Slope Sign Changes (SSC), and Maximum Value (MAX).
   • Autoregressive: AR coefficients (AR1–AR4).
   • Higher-order Crossing: Mean of Amplitude and Derivative Order Crossings (MOAC), Amplitude of Power Order Crossings (APOC), and Standard Deviation of Order Crossings (STDOC).
   • Relative Intensity: RI a mean, RI d mean, RI aad mean.
   • Frequency-domain: Yule–Burg Spectrum (YBS).
3. Preprocessing and Dimensionality Reduction: To reduce redundancy and noise, the study applied:
   • Correlation Filtering: Removal of features with pairwise correlation >95%.
   • Pearson Filtering: Removal of features with absolute Pearson correlation <0.1 relative to class labels.
   • Dimensionality Reduction: Selective application of Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA).
4. Classification: Three classifiers were systematically evaluated across four distinct experimental configurations:
   • Feed-forward Neural Network (NN)
   • k-Nearest Neighbors (KNN)
   • Support Vector Machine (SVM)

Key Results
The study conducted four experiments to isolate the effects of different preprocessing and dimensionality reduction strategies:

• Experiment 1 (Dual Filtering): Applied both Pearson and general correlation filters, reducing features to 8.

  • NN achieved 81% accuracy.
  • KNN and SVM both achieved 78% accuracy.
  • Observation: Performance was limited by the aggressive filtering thresholds.

• Experiment 2 (Pearson Filtering Only): Applied only Pearson filtering, reducing features to 26.

  • NN achieved the study's peak performance of 90% accuracy (and F1 score) using a 4-layer network (256-128-64-32 neurons).
  • KNN and SVM performed poorly (31% and 19%, respectively), attributed to the presence of highly correlated features that negatively impacted distance-based classification.

• Experiment 3 (LDA Reduction): Applied LDA to the 26 Pearson-filtered features, reducing them to 9 components.

  • NN accuracy dropped to 75%.
  • KNN improved to 80%, and SVM to 77%.
  • Observation: LDA successfully mitigated feature redundancy, benefiting distance-based classifiers but slightly degrading NN performance compared to Experiment 2.

• Experiment 4 (PCA Reduction): Applied PCA to the 26 Pearson-filtered features, reducing them to 5 components.

  • NN accuracy dropped to 71%, suggesting information loss due to aggressive reduction.
  • KNN achieved 80%, and SVM achieved 66%.

Significance and Claims
The paper claims that a carefully designed feature-based pipeline can enable accurate single-channel sEMG hand gesture classification without the need for multi-channel arrays. The primary findings indicate:

1. Neural Network Superiority: NNs consistently outperformed KNN and SVM when trained on Pearson-filtered features (Experiment 2), achieving up to 90% accuracy. This suggests that retaining informative time- and frequency-domain features while removing weakly correlated ones is highly effective for NNs.
2. Sensitivity to Redundancy: Distance-based classifiers (KNN, SVM) are highly sensitive to feature redundancy. Dimensionality reduction techniques like LDA and PCA significantly improved their performance by removing correlated features, though they sometimes resulted in information loss for NNs.
3. Feasibility for Embedded Systems: The study demonstrates that combining time and frequency features with Pearson filtering and a compact neural network can achieve competitive accuracy (90%) even with limited temporal and spatial information. This supports the potential for deploying such systems on low-power, embedded platforms (e.g., Raspberry Pi) for cost-effective, real-time human–computer interaction and prosthetic control.

The authors conclude that while single-channel systems involve reduced information compared to multi-channel setups, the proposed methodology offers a viable pathway toward practical, low-power gesture recognition applications.