Comprehensive analysis of supervised learning methods for electrical source imaging.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalography source imaging (ESI) is an ill-posed inverse problem.
• Finding a unique solution requires an additional constraint or prior, which is a significant challenge for existing ESI methods.

Purpose of the Study:

• To explore the application of supervised learning methods for spatio-temporal ESI.
• To train and compare neural networks with non-learning-based methods using synthetic data.

Main Methods:

• Trained three neural networks on synthetic electroencephalography data.
• Used two distinct simulation types based on different brain electrical activity models.
• Quantitatively assessed neural network generalization and training data impact using five metrics.

Main Results:

• Neural networks demonstrated competitive performance compared to non-learning-based ESI methods.
• Performance was evaluated on previously unseen data, showcasing generalization capabilities.
• The design of simulations significantly impacted neural network performance.

Conclusions:

• Supervised learning, particularly neural networks, offers a viable approach for spatio-temporal ESI.
• Appropriately designed simulations are crucial for training effective data-driven ESI models.
• Neural networks can provide competitive solutions for ESI problems, even with novel data.