Estimating Brain Connectivity With Varying-Length Time Lags Using a Recurrent Neural Network.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Estimating brain connectivity aids in understanding information flow and has clinical applications like epilepsy diagnosis.
• Granger causality is a key tool for directional time series analysis but often assumes fixed time lags.
• Brain signal propagation delays are dynamic and vary constantly, posing a challenge for traditional methods.

Purpose of the Study:

• To develop a robust Granger causality estimator capable of handling dynamic and varying time lags in brain connectivity.
• To introduce the Recurrent Neural Network Granger Causality (RNN-GC) estimator for improved multivariate brain connectivity detection.

Main Methods:

• Utilized a gated Recurrent Neural Network (RNN) model, specifically Long Short-Term Memory (LSTM), to process time-series brain signals.
• The LSTM model learns information flow by adaptively updating memory cells to account for variable transmission time lags.
• The RNN-GC estimator was designed to handle arbitrary lengths of transmission time lags in multivariate data.

Main Results:

• The RNN-GC estimator demonstrated superior performance in brain connectivity estimation compared to existing methods.
• Achieved robust modeling of multivariate brain connections with varying-length time lags.
• Showed effectiveness on both simulated data and real-world epileptic electroencephalography (EEG) signals.

Conclusions:

• The RNN-GC method successfully models nonlinear and time-varying lagged information transmission.
• It provides effective directional brain connectivity estimation, even with complex signal dynamics.
• The proposed method shows promise as a reliable tool for clinical brain connection analysis.