Kernel Density Estimation of Electromyographic Signals and Ensemble Learning for Highly Accurate Classification of a Large Set of Hand/Wrist Motions

Parviz Ghaderi, Marjan Nosouhi, Mislav Jordanić, Hamid Reza Marateb, Miguel Ángel Mañanas, Dario FarinaFrontiers in Neuroscience 2022 · DOI

TL;DR

This paper introduces three novel features derived from Kernel Density Estimation (KDE) of surface electromyography (sEMG) signals for improved hand and wrist motion classification. By combining these features with ensemble learning, the researchers achieve state-of-the-art accuracy for both single and simultaneous movements, showing significant promise for advanced myoelectric prosthetics.

What Problem Does This Paper Address?

Controlling prosthetic limbs with myoelectric signals (sEMG) is a critical area for restoring function to amputees. While pattern recognition techniques have advanced, current methods often struggle with classifying a large number of distinct hand and wrist movements, especially in real-time and for simultaneous actions. Existing sEMG features, while useful, can be limited in their ability to capture the nuances of muscle activation patterns, leading to lower classification accuracy, particularly for complex tasks or in individuals with limb loss. This paper aims to bridge this gap by developing more discriminative sEMG features and robust classification strategies.

Methods

The core innovation of this paper lies in the development of three new sEMG features derived from Kernel Density Estimation (KDE). KDE is a non-parametric way to estimate the probability density function of a random variable. The proposed features are:

  • Trimmed Mean of Density (TMD): This feature captures the central tendency of the estimated density, robust to outliers by trimming the extreme 5% of density values.
  • Entropy of Density (ED): This feature quantifies the complexity or uncertainty of the estimated density.
  • Trimmed Mean Absolute Value of Derivative Density (TMAVDD): This feature captures the shape and rate of change of the estimated density, also made robust through trimming.

These features are extracted from each sEMG channel. For single-task classification, Correlation-based Feature Selection (CFS) is used to reduce dimensionality, followed by classification using Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), and Multi-Layer Perceptron (MLP). For simultaneous movement classification, eXtreme Gradient Boosting (XGBoost) is employed. The study also utilizes an ensemble learning approach, combining the outputs of multiple base classifiers to improve overall performance.

Datasets

The study utilizes two publicly available datasets:

  • Ninapro Database V.2: This dataset contains sEMG recordings from 40 intact subjects performing 40 hand and wrist movements. It's used for evaluating conventional (single-task) myoelectric control.
  • Ninapro Database V.3: This dataset includes sEMG recordings from 11 trans-radial amputee subjects performing hand movements. It's used to assess the performance of the proposed features on amputee populations.
  • Ameri's movement dataset: This dataset, recorded by Ameri et al. (2018), contains sEMG signals from 17 intact subjects performing single and simultaneous movements. It's used to evaluate the system's performance in detecting simultaneous actions.

Key Insights and Ideas

  • KDE as a powerful feature extractor: The paper demonstrates that the distribution of sEMG signals, when analyzed through KDE, contains rich information for movement classification, going beyond traditional time-domain features.
  • Ensemble learning is crucial: Combining multiple classifiers (LDA, SVM, MLP) significantly boosts accuracy and robustness compared to individual classifiers, highlighting the benefit of diverse decision-making.
  • New features offer better class separation: The proposed TMD, ED, and TMAVDD features, particularly TMD, show superior ability to distinguish between different hand and wrist movements compared to classic features.
  • Robustness to limb loss: The proposed method shows promising results for amputee subjects, indicating its potential for real-world prosthetic applications.
  • Simultaneous control is achievable: The system effectively classifies simultaneous movements, a critical step towards more natural and intuitive prosthetic control.

Key Results

  • Ninapro Dataset (Single Movements):
    • LDA with the TMD feature achieved an accuracy of 98.99 ± 1.36% for intact subjects and 92.25 ± 9.48% for amputee subjects.
  • Ensemble learning (LDA, SVM, MLP) resulted in an average macro F-score of 98.23 ± 2.02% and micro F-score of 98.32 ± 1.93% for intact subjects.
  • The movement misclassification percentage was 1.75 ± 1.73% for intact and 3.44 ± 2.23% for amputees.
  • Ameri's Dataset (Simultaneous Movements):
    • XGBoost achieved an overall accuracy of 99.71 ± 0.08%.
  • LDA achieved an overall accuracy of 97.85 ± 0.10%.
  • Online Implementation: An online version of the system demonstrated an average accuracy of 82% with a processing time of 63 ms for a 200 ms epoch.

Limitations and Future Work

  • Computational Cost: While the proposed features are promising, the computational complexity of KDE and its derivatives might still be a consideration for extremely low-power embedded systems. The authors suggest vectorization for further optimization.
  • Amputee Ground Truth: The paper acknowledges that obtaining reliable ground truth from amputee subjects can be challenging, as they cannot always produce natural movements.
  • Future Work: The authors suggest further exploration of these features in more complex scenarios, such as continuous control and integration with other sensory modalities. They also highlight the potential for optimizing the online implementation for even faster processing.

Summary updated

Abstract

The performance of myoelectric control highly depends on the features extracted from surface electromyographic (sEMG) signals. We propose three new sEMG features based on the kernel density estimation. The trimmed mean of density (TMD), the entropy of density, and the trimmed mean absolute value of derivative density were computed for each sEMG channel. These features were tested for the classification of single tasks as well as of two tasks concurrently performed. For single tasks, correlation-based feature selection was used, and the features were then classified using linear discriminant analysis (LDA), non-linear support vector machines, and multi-layer perceptron. The eXtreme gradient boosting (XGBoost) classifier was used for the classification of two movements simultaneously performed. The second and third versions of the Ninapro dataset (conventional control) and Ameri's movement dataset (simultaneous control) were used to test the proposed features. For the Ninapro dataset, the overall accuracy of LDA using the TMD feature was 98.99 ± 1.36% and 92.25 ± 9.48% for able-bodied and amputee subjects, respectively. Using ensemble learning of the three classifiers, the average macro and micro-F-score, macro recall, and precision on the validation sets were 98.23 ± 2.02, 98.32 ± 1.93, 98.32 ± 1.93, and 98.88 ± 1.31%, respectively, for the intact subjects. The movement misclassification percentage was 1.75 ± 1.73 and 3.44 ± 2.23 for the intact subjects and amputees. The proposed features were significantly correlated with the movement classes [Generalized Linear Model (GLM); <i>P</i>-value < 0.05]. An accurate online implementation of the proposed algorithm was also presented. For the simultaneous control, the overall accuracy was 99.71 ± 0.08 and 97.85 ± 0.10 for the XGBoost and LDA classifiers, respectively. The proposed features are thus promising for conventional and simultaneous myoelectric control.
BibTeX
@inproceedings{ghaderi2022kernel,
  author = {Parviz Ghaderi and Marjan Nosouhi and Mislav Jordanić and Hamid Reza Marateb and Miguel Ángel Mañanas and Dario Farina},
  title = {{Kernel Density Estimation of Electromyographic Signals and Ensemble Learning for Highly Accurate Classification of a Large Set of Hand/Wrist Motions}},
  booktitle = {Frontiers in Neuroscience},
  year = {2022},
  doi = {10.3389/fnins.2022.796711},
}
Figure