Regression of EEG correlates of simultaneous fMRI signal in epilepsy

Ana Carolina Santa Marta da Silva

Key information

Authors:

Ana Carolina Santa Marta da Silva (Ana Carolina Santa Marta da Silva)

Supervisors:

Patrícia Margarida Piedade Figueiredo (Patrícia Figueiredo); Rodolfo Telo Martins de Abreu (Rodolfo Telo Martins de Abreu)

Published in

11/28/2018

Abstract

The complementary properties of the Electroencephalography (EEG) and the Blood Oxygen Level Dependent (BOLD)-Functional Magnetic Resonance Imaging (fMRI) techniques have been particularly important for mapping epileptic brain networks. We addressed this problem by estimating the EEG correlates -EEG Fingerprint (EFP)- of BOLD-fMRI fluctuations measured at specific distributed epileptic networks previously identified on the fMRI, based on simultaneous EEG-fMRI recordings from a group of five epilepsy patients. The EFPs were based on three power-weighted metrics: Linear Combination of EEG power over frequency bands (LC), Root Mean Squared Frequency (RMSF) and Total Power (TP) - in order to predict the BOLD changes. The resulting EFPs were estimated using different linear regression methods and, their prediction performance was compared by different selection criteria. The considered methods were: Elastic Net Regularization approach including Ridge and Lasso, Least Angle Regression (LAR), Stepwise Regression (SR) and Least-squares (LS). We found that when the predictors from all channels were integrated, the SR yielded the best measures, whereas when the investigation was restricted to each channel, Ridge Regression was chosen due to its favourable advantages dealing with the overfitting effect. In relation to the different metrics, LC with a flexible hemodynamic delay was consistently the best model. In conclusion, this work revealed that SR and Ridge of LC model with no-predefined delay performed best at predicting the BOLD signal of epileptic networks. This methodology may have important applications in the subsequent use of the epileptic EFPs to drive EEG-based Neuro-Feedback (NF) interventions.