Metric validation for detection of delayed and directed coupling.

Area of Science:

• Computational Neuroscience
• Network Science
• Signal Processing

Background:

• The brain's complex neural network underlies all functions, but its dysfunction causes neurological disorders.
• Reconstructing brain network topology from electrophysiology is challenging due to data limitations like noise and sparse sampling.
• Effective connectivity (EC) metrics aim to infer directed neural interactions, but their real-world accuracy is not well understood.

Purpose of the Study:

• To empirically compare the accuracy of common EC metrics in reconstructing known network topologies from simulated electrophysiological data.
• To evaluate metric performance under realistic data constraints, including network size, data length, noise, and coverage.
• To provide guidance on selecting appropriate EC metrics based on data characteristics and research goals.

Main Methods:

• Simulated time series data from Erdős-Rényi networks using a time-delayed vector autoregressive (VAR) model.
• Systematic variation of network size, data length, measurement noise, and network coverage.
• Evaluation of four EC metrics (cross-correlation, Granger causality, mutual information, transfer entropy) using cosine distance and ROC curves.

Main Results:

• Multivariate transfer entropy showed the highest reconstruction accuracy but had long computation times.
• For small networks (<30 nodes), mutual information and Granger causality offered rapid and accurate reconstructions.
• Partial cross-correlation performed well for larger networks with good computational efficiency; zero-lag metrics were ineffective for time-lagged data.

Conclusions:

• The optimal EC metric depends on specific data constraints and network characteristics.
• Multivariate transfer entropy is reliable but computationally intensive; partial cross-correlation is a faster alternative for large networks.
• Time-lagged metrics are essential for accurate network reconstruction, as ignoring temporal delays yields chance-level accuracy.