Classification of Known and Unknown Study Items in a Memory Task Using Single-Trial Event-Related Potentials and...

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Event-related potentials (ERPs) from electroencephalography (EEG) are linked to memory processes.
• Convolutional neural networks (CNNs) can analyze complex neural data at the single-trial level.

Purpose of the Study:

• To assess the feasibility of using EEG-based ERPs as biomarkers for long-term memory classification.
• To compare the performance of different CNN architectures for ERP analysis in memory tasks.

Main Methods:

• Participants (N=25) performed an association memory task while EEG data was recorded.
• Three distinct CNN models, including an EEGNET-based approach, were trained and validated on the ERP data.
• Performance was evaluated using precision, recall, and specificity metrics.

Main Results:

• The EEGNET-based CNN model demonstrated superior performance compared to shallow and deep convolutional models.
• Classification accuracy reached 62% for known items and 66% for unknown items.
• Optimal accuracy involves balancing recall and specificity, influenced by model architecture and dataset size.

Conclusions:

• EEG-derived ERPs, analyzed with CNNs, show promise as biomarkers for long-term memory classification.
• The findings support the integration of ERP and CNNs into online learning tools for memory research.
• This approach can help elucidate the neural mechanisms underlying long-term memorization.