Exploring cognitive workload recognition using CogRepLKNet with EEG-fMRI.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Cognitive Workload Recognition (CWR) faces challenges in integrating multimodal data like EEG and fMRI.
• Heterogeneity of physiological signals complicates unified feature extraction for CWR.

Purpose of the Study:

• To develop a novel deep learning model for accurate multimodal CWR.
• To address the limitations in modeling cross-modal relationships and feature extraction from EEG and fMRI data.

Main Methods:

• Proposed CogRepLKNet, a universal re-parameterizable large-kernel convolutional neural network (CNN).
• Employed parallel perception branches with large- and small-kernel CNNs and adaptive gated attention fusion.
• Utilized input projections for universal feature extraction across physiological signals.

Main Results:

• CogRepLKNet achieved state-of-the-art performance on a self-constructed EEG-fMRI dataset.
• Demonstrated efficient feature integration with reduced computational complexity and fewer training samples compared to transformers.
• Showcased low training complexity and easy portability of the model.

Conclusions:

• CogRepLKNet effectively models cross-modal dynamics for enhanced CWR.
• The model offers a promising solution for multimodal CWR applications.
• The developed approach facilitates advanced brain-computer interfaces and cognitive monitoring.