Deep learning-based EEG analysis: investigating P3 ERP components.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Event-related potentials (ERPs) from electroencephalogram (EEG) analyze neural processing.
• The P3 component, including P3a and P3b, is a key psychophysiological marker for psychiatric disorders.
• Traditional ERP analysis requires averaging, limiting single-subject and single-trial insights due to low signal-to-noise ratio.

Purpose of the Study:

• To investigate a deep learning workflow for enhancing EEG neural signatures of P3 subcomponents at the single-subject and single-trial level.
• To validate the deep learning approach against canonical ERP analysis.

Main Methods:

• A deep learning workflow combining a convolutional neural network (CNN) with an explanation technique (ET) was developed.
• CNN was trained using two strategies: one for cross-subject signatures and another for subject- and trial-specific signatures.
• Saliency representations were generated to visualize enhanced neural signatures.

Main Results:

• Cross-subject saliency maps confirmed P3a and P3b activity within established time windows (350-400 ms frontal, 400-650 ms parietal).
• Single-subject and single-trial saliency maps successfully enhanced P3 signatures at the individual level.
• Standard EEG analysis provided limited or no evident signatures at the single-subject and single-trial level.

Conclusions:

• The CNN+ET workflow effectively enhances EEG neural signatures for P3 subcomponents at single-subject and single-trial levels.
• This method offers a powerful tool for analyzing P3 modulations, providing deeper insights into neural processes linking sensory input, cognition, and behavior.
• The approach holds potential for advancing the understanding and diagnosis of psychiatric disorders.