Deep Convolutional Neural Network Applied to Electroencephalography: Raw Data vs Spectral Features.

Area of Science:

• Neuroscience
• Machine Learning
• Electrophysiology

Background:

• Deep learning has advanced computer vision, prompting exploration in neuroscience for electrophysiological neuroimaging.
• Leveraging deep learning for electroencephalography (EEG) data analysis and decoding is an emerging research area.
• Optimal network architectures and feature spaces for EEG decoding remain open research questions.

Purpose of the Study:

• To compare deep learning performance on raw EEG data versus spectral EEG features.
• To evaluate two distinct deep convolutional neural network architectures for EEG classification.
• To determine the most effective approach for EEG decoding in neuroscience.

Main Methods:

• Compared deep learning models using minimally processed raw EEG data against those using EEG spectral features.
• Utilized two deep convolutional neural network architectures: one tailored for raw data (Putten et al., 2018) and another derived from VGG16 for spectral features (Simonyan and Zisserman, 2015).
• Applied models to classify sex using 24-channel EEG data from 1,574 participants.

Main Results:

• Achieved state-of-the-art classification performance, improving upon existing benchmarks.
• Demonstrated that raw EEG data classification consistently outperformed spectral EEG feature classification across both architectures.
• Observed that the network architecture designed for spectral features showed improved performance when applied to raw data.

Conclusions:

• Raw EEG data, when processed directly by deep learning models, yields superior classification performance compared to spectral features.
• Convolutional neural networks, including those originally designed for spectral features, are highly effective when applied to raw EEG data.
• The findings suggest a paradigm shift towards using raw EEG data in deep learning for enhanced neuroscience research and applications.