Deep learning models as learners for EEG-based functional brain networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Functional brain network (FBN) methods are frequently combined with deep learning (DL) for electroencephalography (EEG) analysis.
• Current approaches typically involve a two-step process: FBN construction for feature extraction followed by DL model analysis.
• Integrating FBN construction directly into DL models could enable end-to-end learning of EEG representations.

Purpose of the Study:

• To investigate whether deep learning models can effectively learn the process of functional brain network construction from EEG data.
• To validate DL models' capacity to learn FBN construction by assessing their ability to reproduce FBN matrices.

Main Methods:

• Utilized deep learning models to learn functional brain network (FBN) matrices derived from electroencephalography (EEG) data.
• Tested seven DL models on two public EEG datasets to learn four representative FBN matrices.
• Assessed model performance using mean squared error (MSE), Pearson correlation coefficient (Corr), and concordance correlation coefficient (CCC).

Main Results:

• Deep learning models demonstrated low MSE and high Corr/CCC for the Coherence network.
• DL models captured the general structure of FBNs but had difficulty modeling specific regions accurately.
• Paired t-tests revealed significant differences (p<0.05) between predicted and actual network topological properties (global efficiency, nodal degree) for most networks.

Conclusions:

• Deep learning models can learn connectivity relationships in certain functional brain networks (FBNs) from EEG data.
• DL models currently struggle to fully capture the intrinsic topological structures of FBNs.
• Hybrid strategies combining traditional FBN methods with deep learning are essential for comprehensive EEG analysis.