Modeling the dynamic brain network representation for autism spectrum disorder diagnosis.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Dynamic functional connectivity analysis offers insights into brain activity during cognitive tasks.
• Understanding spatio-temporal dynamics in brain networks is crucial for elucidating the mechanisms of autism spectrum disorder (ASD).

Purpose of the Study:

• To develop a machine learning approach for classifying neurological disorders, specifically ASD, using interpretable spatio-temporal feature extraction from resting-state fMRI (rs-fMRI).
• To capture and leverage spatio-temporal features within brain networks for improved diagnostic capabilities.

Main Methods:

• rs-fMRI time-series data were transformed into temporal multi-graphs using a sliding window technique.
• Temporal multi-graph clustering was employed to address inconsistencies in the time-series data.
• A graph structure-aware LSTM (GSA-LSTM) model was developed to capture spatio-temporal embeddings and impute incomplete graph data.

Main Results:

• The proposed GSA-LSTM model demonstrated superior performance in classifying brain networks compared to existing state-of-the-art methods on the ABIDE dataset.
• The clustering results derived from the model align with established neuroimaging biomarkers for ASD.

Conclusions:

• The developed dynamic brain network embedding learning approach effectively classifies ASD using rs-fMRI data.
• The GSA-LSTM model provides an interpretable framework for analyzing spatio-temporal brain network dynamics and aids in identifying ASD-related biomarkers.