Two-step clustering-based pipeline for big dynamic functional network connectivity data.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Data Science

Background:

• Dynamic functional network connectivity (dFNC) analyzes time-varying brain network integration using resting-state fMRI (rs-fMRI).
• Conventional dFNC pipelines require extensive computation for clustering and determining the optimal number of brain states.
• Computational demands escalate with larger datasets and longer scan times, limiting accessibility.

Purpose of the Study:

• To develop a computationally efficient dFNC pipeline for analyzing large datasets.
• To reduce the time and computational resources needed for identifying dFNC states.
• To ensure the reproducibility of results across multiple datasets.

Main Methods:

• A novel two-step clustering approach was implemented for dFNC analysis.
• The first step involved sub-sampling data to identify preliminary states at various model orders.
• The second step aggregated these states to determine the optimal cluster number using the elbow criterion, followed by k-means clustering.

Main Results:

• The proposed pipeline achieved over 99% similarity in identified dFNC states compared to conventional methods.
• Analysis time was drastically reduced from 275 minutes to 11 minutes, representing a 25-fold increase in speed.
• The new method demonstrated superior clustering quality (p < 0.001) and reproducible results across four Human Connectome Project datasets.

Conclusions:

• A new, computationally efficient dFNC analysis pipeline has been developed.
• This pipeline enables the analysis of large dFNC datasets without requiring substantial computational resources.
• The method's reproducibility and improved efficiency were validated across multiple datasets.