Predictive assessment of models for dynamic functional connectivity.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Cognitive Neuroscience

Background:

• Dynamic functional connectivity (dFC) models offer a more detailed view of brain function over time compared to static analyses.
• Current dFC models often assume discrete states, but lack consensus on model selection and understanding how assumptions affect state dynamics.

Purpose of the Study:

• To introduce a predictive likelihood approach for assessing and selecting dFC models.
• To evaluate the impact of different modeling assumptions on dFC state estimation.

Main Methods:

• Utilized a predictive likelihood framework to evaluate models based on held-out test data.
• Applied the framework to synthetic data and real-world datasets from EEG (face recognition) and fMRI (resting-state).

Main Results:

• Demonstrated that dynamic modeling approaches better characterize both EEG and fMRI data than static methods.
• Showcased that parameter choices (e.g., window length) and modeling assumptions significantly influence estimated brain states and their interpretation.

Conclusions:

• Dynamic functional connectivity modeling provides superior characterization of brain activity compared to static approaches.
• Caution is advised when interpreting dFC results due to the sensitivity of state estimation to modeling assumptions and parameter choices.