Classification of EEG Signals Using a Multiple Kernel Learning Support Vector Machine

Xiaoou Li, Xun Chen, Yuning Yan, Wenshi Wei, Z. Jane WangSensors 2014 · OpenReview

TL;DR

This paper introduces a Multiple Kernel Learning Support Vector Machine (MKL-SVM) approach for classifying EEG signals during mental and cognitive tasks. By combining multiple kernel functions and leveraging wavelet packet entropy and Granger causality features, the MKL-SVM significantly outperforms traditional single-kernel SVMs, showing promise for advanced brain-computer interfaces, particularly for individuals with cognitive impairments.

What Problem Does This Paper Address?

Brain-Computer Interfaces (BCIs) hold immense potential for communication and control, especially for individuals with motor disabilities. A crucial aspect of BCI systems is accurately interpreting brain activity, often measured by Electroencephalography (EEG). However, EEG signals are inherently noisy and complex, making reliable classification of different mental states or cognitive tasks a significant challenge. Existing methods, while effective to some extent, often struggle with the high dimensionality and non-linear nature of EEG data, leading to suboptimal classification accuracy. This paper aims to improve the accuracy and robustness of EEG signal classification for both mental tasks and cognitive tasks, specifically focusing on identifying individuals with cognitive impairments like stroke patients.

Methods

The proposed approach involves a three-stage process:

  1. Pre-processing: Standard signal processing techniques are applied to enhance the quality of the raw EEG data, aiming to reduce noise and artifacts.
  2. Feature Extraction: Two key feature extraction methods are employed:
  3. Wavelet Packet Entropy (WPE): Used for mental task classification, WPE analyzes the complexity and rhythm information within different frequency bands of the EEG signal, providing a rich representation of brain activity.
  4. Granger Causality: Utilized for cognitive task classification (specifically for stroke patients vs. healthy controls), this method quantifies the directional influence between different brain regions, reflecting functional connectivity.
  5. Classification: A Multiple Kernel Learning Support Vector Machine (MKL-SVM) is the core of the classification strategy. Unlike traditional SVMs that use a single kernel function, MKL-SVM learns an optimal combination of multiple kernels (in this case, polynomial and Radial Basis Function - RBF kernels) by assigning weights to each. This allows the model to adapt better to the complex, multi-faceted nature of EEG data. The MKL-SVM algorithm employed is "SimpleMKL," which uses a gradient descent optimization approach to find these optimal kernel weights. This approach is designed to be efficient and handle multiclass classification problems effectively.

Datasets

The study utilizes two distinct datasets:

  • Mental Task EEG Data: This dataset was originally collected by Keirn and Aunon [9] and is publicly available. It comprises EEG recordings from seven subjects performing five different mental tasks: Baseline, Visual Counting, Mental Letter Composing, Mathematical Multiplication, and Geometric Figure Rotation. The data was sampled at 250 Hz.
  • Cognitive Task EEG Data (N-back Task): This dataset was collected from stroke patients (diagnosed with mild cognitive impairment) and age/sex-matched healthy controls. The cognitive task involved was a verbal N-back task (0-back and 1-back), assessing working memory. EEG data was acquired using an EEG/ERP amplifier system at 256 Hz.

Key Insights and Ideas

  • The Power of Combining Kernels: The central idea is that no single kernel function is universally optimal for EEG classification. By learning to combine multiple kernels (Poly and RBF), MKL-SVM can capture a broader range of signal characteristics, leading to improved performance.
  • Feature Synergy: The combination of WPE for mental tasks and Granger causality for cognitive tasks proved effective. WPE captures signal complexity, while Granger causality highlights functional connectivity, both crucial aspects of brain activity.
  • Individual Differences Matter: The study highlights that the most discriminatory mental tasks can vary significantly between subjects, underscoring the need for adaptive or personalized BCI systems.
  • Stroke Patients Show Reduced Connectivity: The Granger causality analysis revealed a noticeable reduction in effective connectivity between cortical regions in stroke patients compared to healthy controls, providing a neurophysiological basis for their cognitive impairments.
  • MKL-SVM Outperforms Single Kernels: Consistently, the MKL-SVM approach demonstrated superior classification accuracy compared to SVMs using only a single Poly or RBF kernel across various classification scenarios.

Key Results

  • Mental Task Classification:
    • 2-class: Achieved a mean accuracy of 99.20%.
  • 3-class: Mean accuracy of 81.25%.
  • 4-class: Mean accuracy of 76.76%.
  • 5-class: Mean accuracy of 75.25%.
  • MKL-SVM significantly outperformed single-kernel SVMs (Poly and RBF) for all multiclass scenarios.
  • Cognitive Task Classification (Stroke Patients vs. Healthy Controls):
    • 0-back task: Achieved a mean accuracy of 89.24%.
  • 1-back task: Achieved a mean accuracy of 80.33%.
  • MKL-SVM again showed superior performance compared to single-kernel SVMs (Poly and RBF).
  • Computational Cost: The MKL-SVM implementation had an average runtime of 15.39 seconds for training and testing, which is considered practical for BCI applications.

Limitations and Future Work

  • Subject Variability: The study acknowledges that the most easily recognized tasks varied among subjects, suggesting that further research into personalized classification strategies is warranted.
  • Generalizability: While promising, the results are based on specific datasets and tasks. Future work could explore a wider range of mental and cognitive tasks, as well as different subject populations.
  • Real-time Implementation: Although the computational cost is deemed practical, further optimization and validation for real-time BCI applications would be beneficial.
  • Neurophysiological Interpretation: While Granger causality provided insights into connectivity differences in stroke patients, deeper exploration of the neurophysiological underpinnings of the classification results could enhance understanding.
  • Feature Optimization: Investigating other feature extraction techniques or optimizing the current ones could potentially lead to further improvements in classification accuracy.

Summary updated

Abstract

In this study, a multiple kernel learning support vector machine algorithm is proposed for the identification of EEG signals including mental and cognitive tasks, which is a key component in EEG-based brain computer interface (BCI) systems. The presented BCI approach included three stages: (1) a pre-processing step was performed to improve the general signal quality of the EEG; (2) the features were chosen, including wavelet packet entropy and Granger causality, respectively; (3) a multiple kernel learning support vector machine (MKL-SVM) based on a gradient descent optimization algorithm was investigated to classify EEG signals, in which the kernel was defined as a linear combination of polynomial kernels and radial basis function kernels. Experimental results showed that the proposed method provided better classification performance compared with the SVM based on a single kernel. For mental tasks, the average accuracies for 2-class, 3-class, 4-class, and 5-class classifications were 99.20%, 81.25%, 76.76%, and 75.25% respectively. Comparing stroke patients with healthy controls using the proposed algorithm, we achieved the average classification accuracies of 89.24% and 80.33% for 0-back and 1-back tasks respectively. Our results indicate that the proposed approach is promising for implementing human-computer interaction (HCI), especially for mental task classification and identifying suitable brain impairment candidates.
BibTeX
@inproceedings{li2014classification,
  author = {Xiaoou Li and Xun Chen and Yuning Yan and Wenshi Wei and Z. Jane Wang},
  title = {{Classification of EEG Signals Using a Multiple Kernel Learning Support Vector Machine}},
  booktitle = {Sensors},
  year = {2014},
  url = {https://openreview.net/forum?id=e8TAz_gx4eK},
}
Figure