Multiple constraint network classification reveals functional brain networks distinguishing 0-back and 2-back task.

Area of Science:

• Neuroscience
• Cognitive Science
• Artificial Intelligence

Background:

• Working memory is vital for complex cognitive tasks and general intelligence.
• Traditional neuroimaging analyses may miss nonlinear or distributed brain activity patterns.
• Existing methods often rely on linear models, potentially limiting the detection of nuanced brain function.

Purpose of the Study:

• To apply multivariate pattern analysis with a deep learning classifier to whole-brain BOLD activity in children.
• To identify brain activation patterns and functional connectivity distinguishing different working memory loads (0-back vs. 2-back).
• To explore how distinct functional networks contribute to task performance under varying memory demands.

Main Methods:

• Utilized a multiple-constraint deep learning classifier on whole-brain BOLD data from 20 children.
• Trained neural networks to classify task category (0-back vs. 2-back) and activation co-occurrence probability (functional connectivity).
• Employed permutation analyses, model weight examination, and community detection to identify predictive regions and networks.

Main Results:

• Successfully identified global activation patterns and interregional coactivations differentiating memory load conditions.
• Pinpointed brain regions most predictive of memory load and the functional networks integrating them.
• Discovered distinct functional network activation patterns for each memory load, with focused attentional network activation during the 2-back task.

Conclusions:

• Deep learning-based multivariate pattern analysis enhances the detection of brain activity patterns related to working memory.
• Specific functional networks are differentially engaged based on working memory load.
• Findings suggest a more focused attentional network engagement for higher memory demands.