TL;DR
This paper investigates how to effectively classify mental tasks using non-invasive EEG signals by combining Power Spectral Density (PSD) feature extraction with various feature selection techniques. They found that multivariate feature selection methods, particularly the combination of Burg's PSD estimation with Linear Regression and Linear Discriminant Analysis, significantly improve classification accuracy compared to univariate methods.
What Problem Does This Paper Address?
Brain-Computer Interfaces (BCIs) offer a lifeline for individuals with severe disabilities, enabling them to communicate and control devices using their brain signals. A crucial aspect of BCI performance, especially for non-invasive Electroencephalography (EEG), is accurately classifying different mental tasks. However, EEG data is notoriously noisy and high-dimensional, leading to a "curse of dimensionality" where the number of features far exceeds the available training samples. This makes it challenging to build robust classification models. While feature extraction and selection are known to help, there's a need for a systematic study on how different Power Spectral Density (PSD) techniques, when combined with various feature selection methods, perform for mental task classification.
Methods
The authors propose a two-phase approach:
- Feature Extraction: They explore three Power Spectral Density (PSD) estimation methods:
- Welch Method: A non-parametric approach that segments the signal, applies a window, and averages periodograms.
- Burg Method: A parametric approach that models the EEG signal as an autoregressive (AR) process, offering high resolution.
- MUSIC (Multiple Signal Classification): A subspace method that aims to estimate frequency components by analyzing the signal and noise subspaces.
- Feature Selection: They investigate both univariate and multivariate filter-based feature selection techniques:
- Univariate Methods: * Pearson's Correlation: Measures linear relationships between features and class labels.
- Mutual Information (MI): Quantifies the statistical dependency between features and class labels, capturing non-linear relationships.
- Fisher Discriminant Ratio (FDR): A statistical measure of class separability based on feature means and variances.
- Wilcoxon's Ranksum Test: A non-parametric test to compare distributions of two classes.
- Multivariate Methods: * Bhattacharya's Distance: Measures the similarity between probability distributions of different classes.
- Ratio of Scatter Matrices: Evaluates class separability based on within-class and between-class variances.
- Linear Regression: Models the relationship between features and the class label.
- Minimum Redundancy-Maximum Relevance (mRMR): Selects features that are highly relevant to the class label while being minimally redundant among themselves.
The study then compares various combinations of these feature extraction and selection methods using classifiers like Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Support Vector Machines (SVM). They also employ ranking and statistical tests (Friedman test) to determine the best-performing combinations.
Datasets
The study utilizes a publicly available EEG dataset from a previous work by Keirn and Aunon (1990). This dataset includes EEG signals recorded from seven subjects performing five different mental tasks: Baseline (relax), Letter Composing, Mathematical, Visualizing Counting, and Geometric Figure Rotation. The EEG signals were recorded from C3, C4, P3, P4, O1, and O2 electrode positions with A1 and A2 as reference electrodes. Each trial lasted 10 seconds at a sampling rate of 250 Hz. The authors note that Subject 4 was excluded due to incomplete information.
Key Insights and Ideas
- Multivariate Feature Selection Dominance: Multivariate feature selection methods generally outperform univariate methods in improving mental task classification accuracy. This suggests that considering the relationships between features is crucial for effective BCI.
- Burg Method's Efficacy: The Burg method for PSD estimation, a parametric approach, proved to be effective in distinguishing mental tasks, showing notable differences in PSD values across various frequency bands for different tasks.
- FDR as a Strong Univariate Contender: Among the univariate methods, Fisher Discriminant Ratio (FDR) consistently yielded high classification accuracy when combined with PSD approaches.
- Optimal Combination Identified: The combination of the Burg PSD estimation method, Linear Regression (LR) for feature selection, and Linear Discriminant Analysis (LDA) as the classifier achieved the highest classification accuracy.
- MUSIC's Underperformance: The MUSIC method for PSD estimation generally performed worst among the three PSD approaches, regardless of the feature selection method used.
- Feature Redundancy is a Problem: The correlation analysis revealed that some extracted features are highly correlated, highlighting the necessity of feature selection to remove redundancy and improve model performance.
Key Results
The study found that:
- Using feature selection generally improved classification accuracy across all PSD approaches compared to using raw PSD features without selection.
- The combination of Burg PSD + FDR + SVM achieved the highest classification accuracy among univariate feature selection methods.
- The combination of Burg PSD + Linear Regression + LDA achieved the highest classification accuracy among all tested combinations of feature extraction and selection methods.
- Statistical tests (Friedman test) confirmed that multivariate feature selection methods, particularly the combination of LR and Burg, provided statistically significant improvements in classification accuracy compared to many other combinations.
Limitations and Future Work
- Limited Scope of Tasks: The study focused on binary mental task classification. The authors suggest extending this approach to multi-class mental task classification.
- Spectral Density Extraction: The authors propose exploring the extraction of spectral density for different brain frequencies separately in future work.
- Further Investigation: While the study compared various methods, further investigation into the optimal parameters for each method and their interactions could be beneficial.
Summary updated