Home // SIGNAL 2023, The Eighth International Conference on Advances in Signal, Image and Video Processing // View article

A Refined ERR-based Method for Nonlinear System Identification. Application to Epilepsy.

Authors:
Marc Greige
Ahmad Karfoul
Isabelle Merlet
Régine Le Bouquin Jeannès

Keywords: Error Reduction Ratio; Orthogonal Least Squares; proximal optimization; epilepsy; effective connectivity

Abstract:
The goal of this paper is to refine the solution of the Error Reduction Ratio (ERR)-based method for nonlinear system identification in the context of epilepsy. Based on a predefined dictionary, the ERR-based method is composed of two main steps: (i) identifying the most relevant candidates that are required to fit the signal at hand, and (ii) estimating their respective weights in a least squares sense. However, the used candidate selection criterion, which is based on a fixed threshold, often leads to an overestimation of the number of retained candidates. This consequently affects the quality of the system identification. This point is of particular interest in epilepsy especially for the identification of brain networks involved in the seizure onset. To deal with this issue, a refined ERR-based solution is proposed in this paper. It relies on the assumption that a few number of the retained candidates using the ERR-based method are really the most significant ones. This leads to consider a sparse representation of the associated estimated coefficient vector. The well-known Proximal Alternating Linearized Minimization (PALM) is used in this paper to solve the proposed optimization problem. To guarantee good estimation results, the used regularization parameter is, at each iteration, optimally computed using the discrepancy principle. Results on simulated and real iEEG data confirm the efficacy of the proposed method.

Pages: 67 to 71

Copyright: Copyright (c) IARIA, 2023

Publication date: March 13, 2023

Published in: conference

ISSN: 2519-8432

ISBN: 978-1-68558-057-5

Location: Barcelona, Spain

Dates: from March 13, 2023 to March 17, 2023