Virtual EEG-electrodes: Convolutional neural networks as a method for upsampling or restoring channels.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Clinical electroencephalography (EEG) relies on visual assessment, often hampered by reduced electrode counts and artifacts.
• Current interpolation techniques struggle with low electrode density and can produce unreliable distant values.
• A novel method learning cortical electrical field distributions could improve EEG signal reconstruction.

Purpose of the Study:

• To develop and evaluate generative networks for upsampling and restoring EEG signals.
• To compare the performance of generative networks against traditional spherical spline interpolation.
• To assess the impact of data volume on network performance.

Main Methods:

• Convolutional generative networks were trained to upsample from 4/14 channels or restore missing channels to a 21-channel EEG.
• Networks were trained and validated on 5,144 hours of EEG data from 1,385 subjects.
• Performance was evaluated using statistical measures and expert visual assessment by neurophysiologists.

Main Results:

• Generative networks significantly outperformed spherical spline interpolation in EEG signal reconstruction.
• Network-generated data was indistinguishable from real EEG data to expert interpreters.
• Interpolated data was significantly more likely to be identified as artificial.
• Network performance improved with increased subject data, particularly within the 5-100 subject range.

Conclusions:

• Generative neural networks present a viable and effective alternative to spherical spline interpolation for EEG signal restoration and upsampling.
• This AI-driven approach can enhance the quality and reliability of clinical EEG data.
• The findings support the use of deep learning for improving EEG analysis in resource-limited settings.