Recurrent neural networks for reconstructing complex directed brain connectivity.

Area of Science:

• Network Science
• Computational Neuroscience
• Nonlinear Dynamics

Background:

• Granger Causality (GC) is widely used but typically relies on linear models (MVAR).
• Real-world systems, especially biological networks, exhibit significant nonlinear behavior.
• Linear GC methods may inaccurately estimate causal relationships in nonlinear systems.

Purpose of the Study:

• To develop a novel approach for estimating nonlinear, directed network interactions.
• To introduce Echo State Network Granger Causality (ES-GC) for improved causal inference.
• To validate ES-GC's ability to capture nonlinear causal relations.

Main Methods:

• Reformulated Granger Causality using Echo State Networks (ESN), a type of recurrent neural network.
• Simulated multivariate coupling in a network of nonlinearly interacting, noisy Duffing oscillators.
• Applied ES-GC to functional MRI data from 1003 healthy subjects.

Main Results:

• ES-GC demonstrated a superior ability to detect nonlinear causal links compared to existing methods.
• Synthetic validation confirmed ES-GC's advantage in identifying directed interactions.
• Analysis of human brain fMRI data revealed previously unknown between-network interactions.

Conclusions:

• ES-GC is a more effective tool than traditional GC methods for analyzing complex, nonlinear networks.
• The method offers superior performance for biological network analysis, including brain connectivity.
• ES-GC advances the field of causal inference in complex systems.