Providing context: Extracting non-linear and dynamic temporal motifs from brain activity

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Psychiatric Research

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) dynamics are crucial for understanding brain function.
• Current methods often rely on linear approaches for time-resolved functional connectivity (tr-FC), potentially missing complex temporal dynamics.

Purpose of the Study:

• To develop and validate a non-linear deep learning model for analyzing rs-fMRI dynamics.
• To capture multi-scale temporal information in brain activity.
• To identify novel biomarkers for schizophrenia using rs-fMRI.

Main Methods:

• A disentangled variational autoencoder (DSVAE), a generative non-linear deep learning model, was employed.
• The DSVAE factorizes window-specific (context) and timestep-specific (local) information.
• The model's latent space and embeddings were analyzed to compare schizophrenia patients and control subjects.

Main Results:

• Significant differences were found between schizophrenia patients and controls in temporal step distance within the DSVAE's latent space.
• Context embeddings from the DSVAE demonstrated superior separation of patient and control groups compared to standard tr-FC.
• For schizophrenia patients, context embeddings correlated with age and symptom severity, and distinct connectivity patterns were observed in different patient clusters.

Conclusions:

• The DSVAE model captures complementary temporal features beyond standard tr-FC.
• This non-linear approach enhances the analysis of rs-fMRI dynamics.
• The DSVAE shows potential for sensitive detection of psychiatric links and individual characteristics in neuroimaging data.