Spatio-temporal learning and explaining for dynamic functional connectivity analysis: Application to depression.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional connectivity in the brain fluctuates over time.
• Identifying dynamic functional connectivity (dFC) patterns is crucial for understanding brain disorders.
• Major Depressive Disorder (MDD) diagnosis can benefit from advanced neuroimaging analysis.

Purpose of the Study:

• To identify Major Depressive Disorder (MDD) using dynamic functional connectivity (dFC) from resting-state fMRI data.
• To develop tools for early depression diagnosis and enhance understanding of its etiology.
• To classify MDD from healthy controls using a novel spatio-temporal learning framework.

Main Methods:

• Utilized resting-state fMRI data from 178 subjects (89 MDD, 89 controls).
• Developed a stacking neural network model with spatial and temporal encoders for dFC analysis.
• Employed layer-wise relevance propagation (LRP) and attention mechanisms for model interpretability and feature extraction.

Main Results:

• Achieved high classification performance in distinguishing MDD from healthy controls using dFC.
• Identified key functional connections, brain regions, and dynamic states associated with MDD.
• The model successfully uncovered structural and temporal patterns of dFC in depression.

Conclusions:

• The proposed spatio-temporal model effectively classifies MDD based on dFC.
• The study identified specific brain patterns linked to depression.
• Further validation in larger populations and investigation into data preprocessing effects are recommended.