Blind source separation of event-related potentials using a recurrent neural network

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Event-related potentials (ERPs) reflect neural activity linked to specific events.
• Accurate source localization in ERPs is crucial but challenging with conventional methods.
• Existing source separation techniques for ERPs can yield complex, hard-to-interpret results.

Approach:

• Developed a recurrent neural network (RNN) for blind source separation of ERPs.
• The RNN transforms event-related signals into distinct ERP difference waveforms.
• Incorporated L1 regularization for an interpretable, sparse source representation.

Key Points:

• Applied the RNN method to ERPs (MMN, N170, N400, P3) from the ERP CORE database.
• Compared RNN performance against independent component analysis (ICA).
• RNN decomposed ERPs into eleven less noisy, more specific, and distinct sources than ICA.

Conclusions:

• The proposed RNN method effectively decomposes grand-average ERP difference waves.
• RNN-derived sources show reduced ambiguity in amplitude, polarity, and dipole orientation.
• This RNN approach shows promise as a computational model for event-related neural signals.