Robustified principal component analysis for feature selection in EEG signal classification

R. Martin

《International Journal of Systematic Innovation》 7卷5期 (2023/03) Pp. 46-54

https://doi.org/10.6977/IJoSI.202303_7(5).0004

摘要

Feature engineering is an important step in data analysis, especially for machine learning applications. A wide range of feature selection methods are being used in Electroencephalography (EEG) signal processing applications. Principal Component Analysis (PCA) is considered an ideal method for feature selection whenever high dimensional data is obtained, especially in signal processing applications. Following an examination of various EEG signal processing frameworks, PCA emerged as the winner in the battle to reduce dimensionality. Despite its widespread use, it has been found to be ineffective for EEG signal processing problems like epileptic seizure detection due to the nonlinear nature of the signal properties. Traditional methods for solving PCA are insufficient in this case, so suggest a novel technique. In this paper, PCA is explored with an EEG classification model. The proposed work demonstrates how PCA is robustified for an EEG signal processing scenario by applying kernel functions. Statistical features are extracted from EEG data after preprocessing by the Desecrate Wavelet Transform (DWT). Initially, the classical PCA algorithm is applied for feature selection by reducing the dimensionality. Later, the algorithm is robustified by applying a Gaussian kernel in a nonlinear, high-dimensional feature space. In an EEG classification of epileptic seizure detection, the adoption of robustified PCA outperforms conventional PCA in terms of accuracy.

關鍵字

PCA ； Dimensionality Reduction ； Electroencephalography (EEG) ； Feature Engineering ； Signal Processing