A generic framework for embedding human brain function with temporally correlated autoencoder.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Representing human brain function from high-dimensional fMRI data is essential for understanding functional organization.
• Traditional methods like ICA and SDL struggle to effectively capture inter-individual and temporal variability in brain activity.

Purpose of the Study:

• To develop a novel embedding framework for human brain function representation using Transformer models.
• To encode brain function into a compact, comparable latent space for improved analysis of brain activity.

Main Methods:

• Formulated brain functional representation as an embedding problem.
• Proposed a Transformer-based embedding framework to generate dense embedding vectors for brain activities.
• Evaluated the framework on the Human Connectome Project (HCP) task fMRI dataset.

Main Results:

• The learned embeddings demonstrated effectiveness and generalizability in brain state prediction tasks.
• Explored the interpretability of the embeddings from both spatial and temporal perspectives.
• The framework successfully represents brain function regularity and variability in a stereotyped latent space.

Conclusions:

• The proposed Transformer-based embedding framework offers a novel approach to representing human brain function.
• This method provides a general, comparable, and stereotyped latent space for analyzing brain activity across individuals and time.
• The findings offer new insights into the regularity and variability of human brain function.