Abstracts

Rationale: Evaluation of functional connectivity (FC) between intrinsic functional networks is of increasing interest for understanding the epileptic brain. Studies increasingly demonstrate that aberrant interactions between the default mode network (DMN) and "task-positive" resting-state networks play a fundamental role in cognitive deficits in temporal lobe epilepsy (TLE). The majority of previous resting-state functional MRI studies have been based on measures of static connectivity. Accumulating evidence indicates that FC fluctuates dynamically over time, leading to a growing consensus on the importance of adopting a dynamic view of functional connectivity. Despite the promises of dynamic functional connectivity (dFC), however, several challenges remain. Firstly, estimation of functional connectivity as a dynamic quantity leads to an explosion in information complexity. dFC is thus typically evaluated through the mean or variability of fluctuations, whereby a large amount of information potentially informative about normal and epileptic brain connectivity is lost. Secondly, despite the wealth of information provided by dFC, it is unclear how to integrate the vast information in dFC into subject-level prediction for patients with epilepsy. Methods: Interictal resting-state fMRI was performed in 23 healthy controls and 25 patients with temporal lobe epilepsy. A state-space model with Kalman filtering was used to estimate time-varying patterns of functional connectivity between the DMN and several other resting-state networks. Time-frequency analysis was performed on dFC estimates and temporal and spectral features of dFC evaluated. Bootstrap confidence intervals were estimated for each temporal and spectral feature of dFC. Ensemble learning was used to perform subject-level prediction and identification of dFC markers for TLE. Results: A number of temporal and spectral features of dFC, including the spectral rolloff, spectral centroid, spectral spread, spectral skewness, spectral kurtosis, and spectral flatness of dFC between the DMN and other resting-state networks were identified as more informative markers for TLE than traditional connectivity measures. Furthermore, we show that imaging markers that account for subject-level differences in network dynamics attain up to 30% increase in predictive accuracy compared to traditional measures of static connectivity. Conclusions: Our results illuminate many previously unexplored facets of the dynamic properties of functional connectivity between the DMN and other resting-state networks in TLE. Our work provides an interpretable way for integrating the new field of dFC into machine learning and subject-level prediction. Salient information is contained in dynamic functional connectivity, which we show has the capacity to drastically improve the sensitivity of connectomics in biomarker discovery in TLE. Funding: Funding/support for this research was provided by (1) the National Library of Medicine Training Fellowship in Biomedical Informatics, Gulf Coast Consortia for Quantitative Biomedical Sciences (Grant #2T15-LM007093-21); (2) the National Institute of Health (Grant #5T32-CA096520-07); (3) P30-CA016672; (4) The Epilepsy Foundation of America; (5) Baylor College of Medicine Computational and Integrative Biomedical Research Center (CIBR) Seed Grant Awards; (6) Baylor College of Medicine Junior Faculty Seed Funding Program Grant; (7) NIH-NINDS K23 Grant NS044936; (8) The Leff Family Foundation.