Signal identification in ERP data by decorrelated Higher Criticism Thresholding

Emeline Perthame (1) , Ching-Fan Sheu (2) , David Causeur (3, 4)

1 MISTIS - Modelling and Inference of Complex and Structured Stochastic Systems
2 NCKU - National Cheng Kung University
3 IRMAR - Institut de Recherche Mathématique de Rennes
4 LMA2 - Laboratoire de Mathématiques Appliquées Agrocampus

Event-related potentials (ERPs) are intensive recordings of electrical activity along the scalp time-locked to motor, sensory, or cognitive events. A main objective in ERP studies is to select (rare) time points at which (weak) ERP amplitudes (features) are significantly associated with experimental variable of interest. The Higher Criticism Thresholding (HCT), as an optimal signal detection procedure in the " rare-and-weak " paradigm, appears to be ideally suited for identifying ERP features. However, ERPs exhibit complex temporal dependence patterns violating the assumption under which signal identification can be achieved efficiently for HCT. This article first highlights this impact of dependence in terms of instability of signal estimation by HCT. A factor modeling for the covariance in HCT is then introduced to decorrelate test statistics and to restore stability in estimation. The detection boundary under factor-analytic dependence is derived and the phase diagram is correspondingly extended. Using simulations and a real data analysis example, the proposed method is shown to estimate more efficiently the support of signals compared with standard HCT and other HCT approaches based on a shrinkage estimation of the covariance matrix.

Domaines: Statistiques [math.ST]

Emeline Perthame, Ching-Fan Sheu, David Causeur. Signal identification in ERP data by decorrelated Higher Criticism Thresholding. 2016. ⟨hal-01310739⟩