The generalized shrinkage estimator for the analysis of functional connectivity of brain signals

TL;DR

This paper introduces a generalized shrinkage estimator combining parametric and nonparametric spectral density estimates to improve the analysis of functional brain connectivity, validated through simulations and EEG data.

Contribution

The paper presents a novel frequency-specific weighted estimator that adaptively combines parametric and nonparametric spectral estimates for better functional connectivity analysis.

Findings

The estimator outperforms individual methods in simulations.

Effective in identifying frequency-specific brain connectivity.

Validated on EEG data from a visual-motor task.

Abstract

We develop a new statistical method for estimating functional connectivity between neurophysiological signals represented by a multivariate time series. We use partial coherence as the measure of functional connectivity. Partial coherence identifies the frequency bands that drive the direct linear association between any pair of channels. To estimate partial coherence, one would first need an estimate of the spectral density matrix of the multivariate time series. Parametric estimators of the spectral density matrix provide good frequency resolution but could be sensitive when the parametric model is misspecified. Smoothing-based nonparametric estimators are robust to model misspecification and are consistent but may have poor frequency resolution. In this work, we develop the generalized shrinkage estimator, which is a weighted average of a parametric estimator and a nonparametric estimator. The optimal weights are frequency-specific and derived under the quadratic risk criterion so that the estimator, either the parametric estimator or the nonparametric estimator, that performs better at a particular frequency receives heavier weight. We validate the proposed estimator in a simulation study and apply it on electroencephalogram recordings from a visual-motor experiment.