Transformer-aided dynamic causal model for scalable estimation of effective connectivity

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Dynamic Causal Models (DCMs) are crucial for understanding effective connectivity in fMRI data.
• Traditional DCMs face computational limitations, restricting analyses to small brain networks (typically <10 regions).
• Existing faster alternatives like regression DCM (rDCM) have limitations, including linearization and fixed Hemodynamic Response Functions (HRFs).

Purpose of the Study:

• To develop a scalable and accurate method for estimating effective connectivity in large-scale brain networks using fMRI.
• To overcome the computational constraints and inherent limitations of existing DCM approaches.

Main Methods:

• Proposed a novel hybrid approach named Transformer encoder DCM decoder (TREND).
• TREND combines a Transformer encoder with a state-of-the-art physiological DCM (P-DCM) decoder.
• Validated the method through extensive simulations and an empirical fMRI dataset.

Main Results:

• TREND accurately predicts effective connectivity values in networks up to 100 regions.
• Demonstrated significantly reduced computational time compared to P-DCM.
• Showcased superior accuracy and/or speed of TREND against other DCM variants on empirical data.

Conclusions:

• TREND offers a pioneering, scalable solution for effective connectivity analysis in large-scale brain networks.
• The hybrid approach preserves nonlinearities while enhancing computational efficiency.
• TREND extends the applicability of DCMs to complex, whole-brain analyses.