Independent Features Outperform Uncorrelated Approaches in ML Classification

Abstract

Prior EEG research has primarily focused on N-back cognitive task data, utilizing established techniques such as time
frequency spectrum and wavelet-based methods for feature extraction. However, Principal Component Analysis (PCA), 
despite its utility in boosting classifier performance, falls short in capturing nonlinear feature relationships. This study 
proposes a novel approach that integrates Multivariate Empirical Mode Decomposition (MEMD) with Independent 
Component Analysis (ICA) to enhance signal processing and feature extraction. Multivariate empirical mode decomposition 
generates analytic functions from EEG data by decomposing the multichannel EEG into Intrinsic Mode Functions (IMFs), 
from which diverse features are extracted. ICA then further reduces dimensionality, leveraging higher-order statistics to 
pinpoint critical features. The resulting significant, independent features are used to train and test various machine learning 
models, with the k-nearest neighbors algorithm emerging as the most successful, achieving a remarkable 95.27% 
classification accuracy. This approach enhances feature extraction and classification of cognitive task-related EEG data.