Abstract
Previous work has found that simple linear time varying (LTV) models constructed using Ordinary Least Squares regression (OLS) can recreate intracranial EEG (iEEG) signals from complex epileptogenic networks. However, depending on a variety of factors such as the network size, the preprocessing techniques applied, and the dynamics captured within a window, these models may frequently contain unstable eigenvalues, resulting in diverging reconstructions even within short windows of interest. In this paper, we compare four alternative model construction methods to OLS with the aim of both stabilizing the models and improving the signal reconstruction quality. We found that these alternative methods decreased the percentage of diverging reconstructions, but that some methods require careful parameter optimization to prevent over stabilization and converging reconstructions. Through systematic comparison of the effects of the model construction method on both the network connectivity estimates and the model reconstructed iEEG time series, we can now tailor the LTV model construction method for specific applications and use the models generated in downstream analyses to assist with seizure onset zone (SOZ) localization.Clinical Relevance- The use of computational models to model brain network connectivity and reliably reconstruct intracranial neural signals without requiring extensive invasive recordings has the potential to further advance our understanding of epileptogenic networks, improve SOZ localization accuracy and efficiency, and thus improve treatment outcomes for drug-resistant epilepsy patients.