TY - GEN
T1 - On channel selection for EEG-based mental workload classification
AU - Kingphai, Kunjira
AU - Moshfeghi, Yashar
N1 - Copyright © 2024 Springer-Verlag. This version of the article has been accepted for publication, after peer review (when applicable) and is subject to Springer Nature’s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at https://doi.org/10.1007/978-3-031-53966-4_30
PY - 2024/2/15
Y1 - 2024/2/15
N2 - Electroencephalogram (EEG) is a non-invasive technology with high temporal resolution, widely used in Brain-Computer Interfaces (BCIs) for mental workload (MWL) classification. However, numerous EEG channels in current devices can make them bulky, uncomfortable, and time-consuming to operate in real-life scenarios. A Riemannian geometry approach has gained attention for channel selection to address this issue. In particular, Riemannian geometry employs covariance matrices of EEG signals to identify the optimal set of EEG channels, given a specific covariance estimator and desired channel number. However, previous studies have not thoroughly assessed the limitations of various covariance estimators, which may influence the analysis results. In this study, we aim to investigate the impact of different covariance estimators, namely Empirical Covariance (EC), Shrunk Covariance (SC), Ledoit-Wolf (LW), and Oracle Approximating Shrinkage (OAS), along with the influence of channel numbers on the process of EEG channel selection. We also examine the performance of selected channels using diverse deep learning models, namely Stacked Gated Recurrent Unit (GRU), Bidirectional Gated Recurrent Unit (BGRU), and BGRU-GRU models, using a publicly available MWL EEG dataset. Our findings show that although no universally optimal channel number exists, employing as few as four channels can achieve an accuracy of 0.940 (±0.036), enhancing practicality for real-world applications. In addition, we discover that the BGRU model, when combined with OAS covariance estimators and a 32-channel configuration, demonstrates superior performance in MWL classification tasks compared to other estimator combinations. Indeed, this study provides insights into the effectiveness of various covariance estimators and the optimal channel subsets for highly accurate MWL classification. These findings can potentially advance the development of EEG-based BCI applications.
AB - Electroencephalogram (EEG) is a non-invasive technology with high temporal resolution, widely used in Brain-Computer Interfaces (BCIs) for mental workload (MWL) classification. However, numerous EEG channels in current devices can make them bulky, uncomfortable, and time-consuming to operate in real-life scenarios. A Riemannian geometry approach has gained attention for channel selection to address this issue. In particular, Riemannian geometry employs covariance matrices of EEG signals to identify the optimal set of EEG channels, given a specific covariance estimator and desired channel number. However, previous studies have not thoroughly assessed the limitations of various covariance estimators, which may influence the analysis results. In this study, we aim to investigate the impact of different covariance estimators, namely Empirical Covariance (EC), Shrunk Covariance (SC), Ledoit-Wolf (LW), and Oracle Approximating Shrinkage (OAS), along with the influence of channel numbers on the process of EEG channel selection. We also examine the performance of selected channels using diverse deep learning models, namely Stacked Gated Recurrent Unit (GRU), Bidirectional Gated Recurrent Unit (BGRU), and BGRU-GRU models, using a publicly available MWL EEG dataset. Our findings show that although no universally optimal channel number exists, employing as few as four channels can achieve an accuracy of 0.940 (±0.036), enhancing practicality for real-world applications. In addition, we discover that the BGRU model, when combined with OAS covariance estimators and a 32-channel configuration, demonstrates superior performance in MWL classification tasks compared to other estimator combinations. Indeed, this study provides insights into the effectiveness of various covariance estimators and the optimal channel subsets for highly accurate MWL classification. These findings can potentially advance the development of EEG-based BCI applications.
KW - mental workload classification
KW - machine learning
KW - deep learning
KW - channel selection
KW - covariance estimator
U2 - 10.1007/978-3-031-53966-4_30
DO - 10.1007/978-3-031-53966-4_30
M3 - Conference contribution book
SN - 9783031539657
T3 - Lecture Notes in Computer Science
SP - 403
EP - 417
BT - Machine Learning, Optimization, and Data Science
A2 - Nicosia, Giuseppe
A2 - Ojha, Varun
A2 - La Malfa, Emanuele
A2 - La Malfa, Gabriele
A2 - Pardalos, Panos M.
A2 - Umeton, Renato
PB - Springer-Verlag
CY - Cham
T2 - Third Symposium on Artificial Intelligence and Neuroscience
Y2 - 22 September 2023 through 26 September 2023
ER -