The More, the Better? Evaluating the Role of EEG Preprocessing for Deep Learning Applications

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning (DL) shows promise for electroencephalography (EEG) analysis, surpassing traditional methods.
• Effective DL model performance hinges on appropriate data preprocessing, but optimal strategies for EEG remain unclear.
• Lack of consensus on EEG preprocessing for DL leads to uncertainty in research.

Purpose of the Study:

• To comprehensively investigate the impact of varying EEG preprocessing levels on deep learning model performance.
• To establish guidelines for EEG data preprocessing in deep learning applications.
• To evaluate the influence of artifact removal techniques on DL model generalizability.

Main Methods:

• Evaluated diverse preprocessing pipelines, from raw to complex artifact removal, on six EEG classification tasks.
• Assessed four established deep learning architectures across these tasks.
• Trained and analyzed 4800 distinct models to compare preprocessing pipeline effects.

Main Results:

• Models trained on raw EEG data consistently exhibited the poorest performance across all tasks.
• Minimal preprocessing pipelines, particularly those without automated artifact handling, generally led to superior model outcomes.
• Statistical differences in performance were observed based on preprocessing strategies at both intra-task and inter-task levels.

Conclusions:

• EEG data preprocessing significantly influences deep learning model performance and generalizability.
• Raw EEG data is suboptimal for training deep learning models.
• Minimal preprocessing pipelines may be more beneficial for deep learning in EEG, suggesting artifacts impact network performance.