I see artifacts: ICA-based EEG artifact removal does not improve deep network decoding across three BCI tasks.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Automated preprocessing of electroencephalography (EEG) data is crucial for transparency and reproducibility.
• Independent component (IC)-based methods are popular for identifying and removing artifacts in EEG.
• The efficacy of IC-based noise rejection in multivariate scenarios like neural network decoding remains unclear.

Purpose of the Study:

• To investigate the impact of IC-based noise rejection on neural network classifier-based decoding of EEG data.
• To compare different IC decomposition and component rejection strategies across diverse EEG datasets.
• To evaluate the trade-off between performance gains and computational cost of IC-based preprocessing.

Main Methods:

• Applied two IC decomposition methods (Infomax, AMICA) and three rejection strategies (none, ICLabel, MARA).
• Utilized three EEG datasets: motor imagery, long-term memory, and visual memory.
• Cross-validated processed data with three neural network architectures (two CNNs, one LSTM).

Main Results:

• IC-based noise rejection provided at best minor benefits for EEG decoding.
• Component-rejected data did not consistently outperform data without rejection.
• Significant computational resources are required for IC analysis, questioning the cost-benefit ratio.

Conclusions:

• The utility of IC-based noise rejection in neural network-based EEG decoding is questionable.
• Current IC-based artifact removal methods may not justify the computational expense for decoding tasks.
• Further research may be needed to optimize automated EEG preprocessing for multivariate analyses.