How reliable are MEG resting-state connectivity metrics?

Colclough GL, Woolrich MW, Tewarie PK, Brookes MJ, Quinn AJ, Smith SM. · 2016

Read How reliable are MEG resting-state connectivity metrics? by Colclough GL, Woolrich MW, Tewarie PK, Brookes MJ, Quinn AJ, Smith SM. (2016) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5056955.txt.

Introduction

G. L. Colclough, M. W. Woolrich, P. K. Tewarie, M. J. Brookes, A. J. Quinn, and S. M. Smith test brain connectivity measures for reliability in NeuroImage in 2016. MEG records the magnetic fields of brain activity with fine timing, and many methods turn those signals into resting-state networks, with little guidance on which to trust. Using Human Connectome Project data, the team ran twelve network methods and judged them by repeatability within a person and across a group.

A confound dominated the results: magnetic signals leak between nearby sources, which can fake connections and inflate consistency scores. Phase- and coherence-based methods that resist the leak repeated poorly over time. The most consistent methods were correlation of signal amplitude envelopes and partial correlation; the orthogonalized envelope method was the authors' pick. Their advice is to choose measures built to withstand leakage before trusting any resting-state network.

Catalog record

CREATORColclough GL, Woolrich MW, Tewarie PK, Brookes MJ, Quinn AJ, Smith SM.
YEAR2016
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC5056955

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

How reliable are MEG resting-state connectivity metrics? By G.L. Colclough, M.W. Woolrich, P.K. Tewarie, M.J. Brookes, A.J. Quinn, S.M. Smith Published in: Neuroimage DOI: 10.1016/j.neuroimage.2016.05.070 ABSTRACT MEG offers dynamic and spectral resolution for resting-state connectivity which is unavailable in fMRI. However, there are a wide range of available network estimation methods for MEG, and little in the way of existing guidance on which ones to employ. In this technical note, we investigate the extent to which many popular measures of stationary connectivity are suitable for use in resting-state MEG, localising magnetic sources with a scalar beamformer. We use as empirical criteria that network measures for individual subjects should be repeatable, and that group-level connectivity estimation shows good reproducibility. Using publically-available data from the Human Connectome Project, we test the reliability of 12 network estimation techniques against these criteria. We find that the impact of magnetic field spread or spatial leakage artefact is profound, creates a major confound for many connectivity measures, and can artificially inflate measures of consistency. Among those robust to this effect, we find poor test-retest reliability in phase- or coherence-based metrics such as the phase lag index or the imaginary part of coherency. The most consistent methods for stationary connectivity estimation over all of our tests are simple amplitude envelope correlation and partial correlation measures. INTRODUCTION Resting-state connectivity estimation with electrophysiology is an important tool for studying intrinsic brain activity. It is able to act in concert with resting-state functional magnetic resonance imaging (fMRI), both by deconstructing the oscillation-specific functional origin of resting-state networks (Brookes et al., 2011, Marzetti et al., 2013), and as a feasible tool for assessing the short-timescale fluctuations of connectivity at rest (Baker et al., 2014, de Pasquale et al., 2010, de Pasquale et al., 2015). It acts as a complementary modality to fMRI and electrocorticograms for analyses of cortical communication patterns (de Pasquale et al., 2015), as well as of the fundamental differences in connectivity between healthy and diseased populations (Stam and van Straaten, 2012, Stam, 2014, van Straaten and Stam, 2013). There are some key challenges associated with connectivity modelling in magnetoencephalography (MEG). Firstly, the activity of the cortical sources responsible for the sensor recordings must be estimated (connectivity estimation between sensors being hard to interpret; Schoffelen2009, Schoffelen2009). Source estimates are highly spatially correlated, and this spatial leakage of inferred sources into their local neighbourhood can create the semblance of connectivity between regions that in truth just share components of the same sensor signal (Palva and Palva, 2012). Secondly, the relatively low rank of electrophysiological data places a limit on the number of independent network nodes that can be sensibly specified in a connectivity model (there is little merit in estimating a network matrix over a 2 mm source grid, for example). We can, however, attempt to describe as many of the dominant cortical sources as our data allow using sensible regions of interest (ROIs). Thirdly, the low signal-to-noise ratio (SNR) of MEG and EEG data creates noisy connectivity estimates: for example, there is less consensus about the spatial characterisation of conventional resting-state networks as expressed in MEG than there is in fMRI. Lastly (and this challenge is shared in fMRI modelling), it may be desired to estimate just the direct functional connections within the cortical network, or to assign directionality of information flow between nodes. Methods for achieving these goals, for example by partialling out data (Marrelec et al., 2006) or using temporal precedence as an indicator of causality (Granger, 1969, Baccala and Sameshima, 2001), are noisier to estimate, and the extent to which they show repeatable or reliable behaviour in MEG is poorly understood. A large number of connectivity estimation methods are available to resting-state researchers. These may assess spectral coherence, the presence of a consistent phase delay, or the amplitude coupling of oscillatory patterns. Some of these methods estimate directionality, and some direct network connections. Not all authors use metrics which are insensitive to spatial leakage confounds, even after recent work highlighting the associated problems (Palva and Palva, 2012, Colclough et al., 2015, Brookes et al., 2012, Hipp et al., 2012, Maldjian et al., 2014). Despite this wide range of practice, the reliability and consistency of these network measures have not been fully compared. As a result, many researchers may be picking analysis techniques without an understanding of the reproducibility of their results. We seek to provide guidance to the researcher who is setting out on a resting-state analysis as to which network estimation method to choose. We believe a key property that any network connectivity measure must satisfy is the repeatability and consistency in estimation of both the single-subject result and the group-level structure. Consistency of estimation is a measure of the noisiness of a network metric, given a set of resting-state scan protocols and source reconstruction methods. Indeed, the sensitivity of inference on the mean effect, by the reduction of noisy sources of variability, can be a good surrogate test in the absence of a known ground truth, even in the presence of smaller (and potentially informative) sources of variation in the network matrices (Smith et al., 2005). Knowledge of estimation consistency is also clinically relevant, as a key requirement in the development of normative population distributions. An understanding of the run-to-run variability of network estimation is a crucial first step when moving onward from group comparison studies to the classification or characterisation of new subjects. In this article, we compare the consistency and reproducibility of whole-brain network estimation between 39 functionally-defined cortical ROIs both at the group-level, and within and between individual subjects for 12 commonly-employed network estimation metrics. We include in our analysis connectivity measures which are susceptible to spatial leakage confounds, not because they are necessarily to be recommended but because they are still used in the field. We will explore the extent to which the false connectivity induced by these confounds can artificially increase the consistency of network estimation. We concentrate on stationary connectivity, rather than characterising the evolution of networks over time in each subject. We perform our analysis on the resting-state MEG data provided by the Human Connectome Project (HCP), for a number of reasons. The data are freely available, have passed through a stringent pre-processing pipeline, include a large number of young, healthy subjects (61 with resting scans in the MEG2 release), and include three repeated sessions of resting-state recording for each participant. As an alternative to the range of methods for connectivity estimation that are robust to spatial leakage artefacts, there are some source reconstruction approaches that are designed to minimise the cross-talk between specific cortical locations (for example, Hui et al. (2010), Hauk and Stenroos (2014), Dalal et al. (2006)). These methods require the multi-source reconstruction to have fewer dipoles than the rank of the sensor data: each ROI would have to be described by a single principle component of its lead fields or a representative voxel. The relative benefits of performing the dimensionality reduction down to the number of ROIs either before or after the source reconstruction (as we do in this paper) remain to be explored. Other authors have essayed similar studies, but without the range of metrics presented here, or without the focus on

Read the complete text