Decoding study-independent mind-wandering from EEG using convolutional neural networks.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Mind-wandering is a common cognitive state where attention shifts from external tasks to internal thoughts.
• Accurately detecting mind-wandering is crucial for understanding attention and cognitive control.
• Existing methods for mind-wandering detection often lack generalizability across different studies and participants.

Purpose of the Study:

• To develop and evaluate study-independent electroencephalography (EEG) classifiers for tracking mind-wandering.
• To investigate the effectiveness of different EEG data transformations as input for machine learning models.
• To assess the generalizability of trained classifiers across participants and independent studies.

Main Methods:

• EEG data was transformed into band-frequency power, single-trial ERP (stERP) patterns, and channel connectivity matrices.
• Convolutional neural networks (CNNs) were trained on each input type for mind-wandering classification.
• A meta-learner was trained using the outputs of the CNN models, with leave-N-participant-out cross-validation and independent study testing for generalizability assessment.

Main Results:

• Classifiers demonstrated limited generalizability across participants and tasks.
• The meta-learner utilizing stERP patterns achieved superior performance compared to other state-of-the-art neural networks.
• Analysis of the meta-learner's input mapping revealed the importance of specific EEG channels for mind-wandering detection.

Conclusions:

• Training study-independent mind-wandering classifiers using EEG is a challenging but important endeavor.
• The proposed stacking neural network design facilitates the inspection of channel importance and feature maps.
• Single-trial ERP patterns show potential as a robust feature for improving mind-wandering detection accuracy.