Estimation of brain connectivity through Artificial Neural Networks

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Estimating brain connectivity from electroencephalographic signals (EEG) is crucial for understanding brain function.
• Multivariate Autoregressive (MVAR) models are effective for EEG connectivity analysis.
• Artificial Neural Networks (ANNs) can serve as MVAR models but struggle with limited data due to regressor collinearity.

Purpose of the Study:

• To evaluate the efficacy of ANNs trained with Stochastic Gradient Descent-L1 (SGD-L1) for brain connectivity estimation using EEG data with limited samples.
• To address the limitations of traditional MVAR parameter estimation in low-data scenarios.

Main Methods:

• Utilized ANNs as MVAR models for brain connectivity estimation.
• Implemented the SGD-L1 algorithm for efficient training of ANNs with L1-norm regularization.
• Tested the approach on both surrogate and real EEG datasets with limited sample sizes.

Main Results:

• ANNs trained with SGD-L1 demonstrated accurate MVAR parameter estimation even with few data samples.
• The SGD-L1 algorithm effectively mitigated issues arising from regressor collinearity in low-data conditions.
• The approach showed robust performance on both simulated and real-world EEG data.

Conclusions:

• ANNs combined with SGD-L1 offer a viable solution for accurate brain connectivity estimation from EEG, particularly when data is scarce.
• This method overcomes limitations of traditional approaches in low-sample regimes.
• The findings suggest a promising direction for advancing EEG-based brain connectivity research.