Robust learning from corrupted EEG with dynamic spatial filtering.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Machine learning models for electroencephalography (EEG) often struggle with noisy data and missing channels, especially in non-laboratory settings.
• Sparse EEG montages (1-6 channels) common in mobile or consumer devices exacerbate these challenges.
• Existing methods for handling missing EEG data are often impractical for resource-limited applications.

Purpose of the Study:

• To develop a robust method for analyzing EEG data with missing channels, particularly for sparse montages in real-world applications.
• To introduce Dynamic Spatial Filtering (DSF) as a novel attention module for improving EEG analysis robustness.
• To evaluate the performance of DSF against traditional methods under various channel corruption scenarios.

Main Methods:

• Proposed Dynamic Spatial Filtering (DSF), a multi-head attention module integrated into neural networks.
• DSF learns to focus on valid EEG channels and disregard corrupted ones.
• Tested DSF on extensive public EEG datasets with simulated corruption and a private dataset with natural corruption from at-home mobile EEG recordings.

Main Results:

• DSF achieved performance comparable to baseline models on clean EEG data.
• DSF significantly outperformed baseline models by up to 29.4% accuracy in the presence of substantial channel corruption.
• DSF provides interpretable outputs, allowing real-time monitoring of channel importance.

Conclusions:

• Dynamic Spatial Filtering (DSF) offers a robust solution for EEG analysis in challenging, real-world conditions with channel corruption.
• DSF enhances the reliability and accuracy of machine learning models using sparse EEG data.
• The interpretability of DSF facilitates practical application in mobile and wearable EEG devices.