Related Experiment Video
Updated: Sep 16, 2025
06:40
Microstate and Omega Complexity Analyses of the Resting-state Electroencephalography
Published on: June 15, 2018
10.3K
How EEG preprocessing shapes decoding performance
Roman Kessler1, Alexander Enge2,3, Michael A Skeide2
1Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany. rkesslerx@gmail.com.
Communications Biology
|July 10, 2025
Summary
Electroencephalography (EEG) preprocessing significantly impacts classification performance. Artifact correction reduced decoding, while specific filtering and detrending improved it, highlighting the need for careful preprocessing selection.
Area of Science:
- Neuroscience
- Machine Learning
- Signal Processing
Background:
- Electroencephalography (EEG) preprocessing methods vary widely across studies.
- The impact of these preprocessing choices on classification performance is not well understood.
Purpose of the Study:
- To investigate how different EEG preprocessing steps affect decoding performance.
- To identify which preprocessing parameters most influence classification accuracy.
Main Methods:
- Systematic variation of preprocessing steps (filtering, referencing, baseline correction, detrending, artifact correction) using MNE-Python.
- Trial-wise binary classification using neural networks (EEGNet) and time-resolved logistic regressions on the ERP CORE dataset.
- Analysis of seven experiments with 40 participants.
Main Results:
- Preprocessing choices significantly influenced decoding performance.
- Artifact correction steps generally reduced performance, while higher high-pass filter cutoffs increased it.
- Specific steps like baseline correction (EEGNet) and linear detrending (logistic regression) improved performance, with other effects being experiment-specific.
Conclusions:
- Careful selection of EEG preprocessing steps is crucial for reliable decoding.
- While artifact correction may inflate performance, it can compromise interpretability and model validity by exploiting noise.
- Preprocessing choices should be optimized based on the specific experiment and event-related potential component.
