[Paper Review] Feature Analysis for Classification of Physical Actions using surface EMG Data

Anish C. Turlapaty, Balakrishna Gokaraju|arXiv (Cornell University)|Nov 9, 2018

Muscle activation and electromyography studies36 references41 citations

TL;DR

This paper proposes a novel feature engineering framework for classifying physical actions using multi-channel surface EMG signals, introducing modified spectral moment ratios/products and inter-channel correlation features to enhance classification accuracy. Using a probabilistic neural network (PNN) with 37 selected features, the method achieves 92.75% accuracy and 0.924 Cohen’s kappa, outperforming baseline models and demonstrating strong performance comparable to SVM with reduced computational cost.

ABSTRACT

Based on recent health statistics, there are several thousands of people with limb disability and gait disorders that require a medical assistance. A robot assisted rehabilitation therapy can help them recover and return to a normal life. In this scenario, a successful methodology is to use the EMG signal based information to control the support robotics. For this mechanism to function properly, the EMG signal from the muscles has to be sensed and then the biological motor intention has to be decoded and finally the resulting information has to be communicated to the controller of the robot. An accurate detection of the motor intention requires a pattern recognition based categorical identification. Hence in this paper, we propose an improved classification framework by identification of the relevant features that drive the pattern recognition algorithm. Major contributions include a set of modified spectral moment based features and another relevant inter-channel correlation feature that contribute to an improved classification performance. Next, we conducted a sensitivity analysis of the classification algorithm to different EMG channels. Finally, the classifier performance is compared to that of the other state-of the art algorithms

Motivation & Objective

• To improve the classification accuracy of physical actions from multi-channel sEMG signals using advanced feature engineering.
• To identify and integrate novel features—specifically modified spectral moment ratios/products and inter-channel correlation features—that enhance pattern recognition performance.
• To evaluate the sensitivity of classification performance to different EMG channels and determine the most relevant ones for action classification.
• To compare the performance of the PNN classifier against state-of-the-art algorithms like multi-class SVM in the context of sEMG-based action recognition.
• To establish a robust, low-computation framework suitable for real-time control of upper-limb exoskeletons in rehabilitation robotics.

Proposed method

• Extracted 276 features from eight sEMG channels, including time-domain statistics, inter-channel cross-correlation, spectral band powers (via Burg algorithm), and local binary pattern (LBP) features.
• Proposed modified spectral moment ratios and products as enhanced frequency-domain features to better capture discriminative patterns in EMG signals.
• Introduced an inter-channel correlation feature (ICS) to model cross-channel dependencies, improving classification robustness.
• Applied Sequential Forward Selection (SFS) to identify the most relevant 37 features from the 276 candidates, optimizing for classification performance.
• Employed a Probabilistic Neural Network (PNN) as the classifier, trained and validated using 10-fold cross-validation on a multi-category action dataset.
• Compared PNN performance against multi-class SVM with polynomial kernel using the same feature subset to validate robustness and efficiency.

Experimental results

Research questions

• RQ1 Can modified spectral moment ratios and products significantly improve the classification of physical actions from sEMG signals compared to conventional features?
• RQ2 How does the inclusion of inter-channel correlation features affect the performance of sEMG-based action classification?
• RQ3 Which EMG channels contribute most significantly to accurate classification of physical actions?
• RQ4 Does the PNN classifier achieve comparable or better performance than kernel-based SVMs while requiring less computational overhead?
• RQ5 To what extent does the proposed feature set reduce misclassification, particularly between similar actions like front kick and side kick?

Key findings

• The inclusion of the proposed modified spectral moment ratios and products (LMF) improved the Cohen’s Kappa coefficient by 9% compared to the baseline feature set.
• Adding inter-channel correlation features (ICS) alone increased the Kappa score by 4%, demonstrating their value in capturing cross-channel dynamics.
• The combined use of LMF and ICS features led to an overall 11.75% improvement in Kappa, achieving a final Kappa of 0.924 and accuracy of 92.75%.
• The confusion matrix revealed that normal actions were classified with high accuracy (e.g., 58–60 correct out of 60 trials), while aggressive actions showed slightly higher misclassification, particularly between front kick (class 19) and side kick (class 12), with 16 misclassified patterns.
• The PNN classifier achieved 92.75% accuracy and 0.924 Kappa, closely matching the SVM’s 91.5% accuracy and 0.91 Kappa, but with lower computational cost.
• The upper limb EMG channels were found to be most relevant for action classification, indicating their dominance in encoding motor intent for upper-limb movements.