This methods study benchmarks common effective connectivity metrics on simulated time-delayed networks and finds multivariate transfer entropy most accurate but computationally intensive, while zero-lag metrics fail for time-lagged interactions and simpler metrics (Granger causality, mutual…

Reliable EC metric selection for noisy, sparsely sampled human electrophysiology improves reconstruction of directed brain networks, which is a necessary step toward robust electrophysiology-based biomarkers and circuit-targeted therapies (e.g., DBS optimization) in Parkinson's disease.

The brain functions as a complex network of billions of interconnected neurons, coordinating processes from basic reflexes to high-level cognition. Dysfunction in these networks contribute to neurological and psychiatric disorders, including epilepsy, depression, and Parkinson's disease. Understanding these network alterations is essential for developing effective therapies. However, reconstructing network topology from human electrophysiology data is challenging due to sparse spatial sampling, measurement noise, and variable time delays in interregional communication. Effective connectivity (EC) metrics have been developed to infer directed neural interactions, but their accuracy under real-world data constraints remain unclear. This study empirically compares the ability of common EC metrics to reconstruct relationships between simulated time series with known temporal relationships and network topologies in the presence of data limitations common to human electrophysiology data. By utilizing networks and temporal relationships that are mathematically simple, this framework provides broad conceptual backing to understand the reliability of EC metrics and establishes groundwork upon which more complex spatial and temporal relationships between time series can be evaluated.&#xD;Approach: We generated Erdős-Rényi networks and simulated time series using a time-delayed vector autoregressive (VAR) model. We systematically varied network size, data length, measurement noise, and network coverage. Variations of four commonly used EC metrics, cross-correlation, Granger causality, mutual information, and transfer entropy, were evaluated for reconstruction accuracy using cosine distance, as well as receiver operating characteristic (ROC) curves, to compare estimated and true coupling matrices. &#xD;Main Results: Multivariate transfer entropy demonstrated the highest accuracy across various conditions but required significantly longer computation times. For small networks (<30 nodes), mutual information and Granger causality rapidly and accurately reconstructed networks. For larger networks, partial cross-correlation performed well with good computational efficiency. Notably, zero-lag metrics perform no better than chance for time-lagged time series relationships in nearly all conditions.&#xD;Significance: The choice of an EC metric should consider specific data constraints. While multivariate transfer entropy is the most reliable across conditions, its long runtime limits its practical application. For large networks, partial cross-correlation offers a faster and reasonably accurate alternative. Granger causality and mutual information are effective for small networks. Critically, time-lagged metrics are essential for accurate network reconstructions, as failing to account for time delays leads to reconstructions no more accurate than random network models.