|
| 1 | +--- |
| 2 | +layout: default |
| 3 | +title: Chapter-2.-Introduction |
| 4 | +long_title: Chapter-2.-Introduction |
| 5 | +parent: SIFT |
| 6 | +grand_parent: Plugins |
| 7 | +--- |
| 8 | +Mapping the structural and active functional properties of brain |
| 9 | +networks is a key goal of basic and clinical neuroscience and medicine. |
| 10 | +The novelty and importance of this transformative research was recently |
| 11 | +emphasized by the U.S. National Institute of Health in their 2010 |
| 12 | +announcement for the Human Connectome Project: |
| 13 | + |
| 14 | +> *Knowledge of human brain connectivity will transform human |
| 15 | +> neuroscience by providing not only a qualitatively novel class of |
| 16 | +> data, but also by providing the basic framework necessary to |
| 17 | +> synthesize diverse data and, ultimately, elucidate how our brains work |
| 18 | +> in health, illness, youth, and old age.* |
| 19 | +
|
| 20 | +The study of human brain connectivity generally falls under one or more |
| 21 | +of three categories: *structural*, *functional*, and *effective* |
| 22 | +(Bullmore and Sporns, 2009). |
| 23 | + |
| 24 | +*Structural connectivity* denotes networks of anatomical (e.g., axonal) |
| 25 | +links. Here the primary goal is to understand what brain structures are |
| 26 | +*capable* of influencing each other via direct or indirect axonal |
| 27 | +connections. This might be studied *in vivo* using invasive axonal |
| 28 | +labeling techniques or noninvasive MRI-based diffusion weighted imaging |
| 29 | +(DWI/DTI) methods. |
| 30 | + |
| 31 | +*Functional connectivity* denotes (symmetrical) correlations in activity |
| 32 | +between brain regions during information processing. Here the primary |
| 33 | +goal is to understand what regions are functionally related through |
| 34 | +correlations in their activity, as measured by some imaging technique. A |
| 35 | +popular form of functional connectivity analysis using functional |
| 36 | +magnetic resonance imaging (fMRI) has been to compute the pairwise |
| 37 | +correlation (or partial correlation) in BOLD activity for a large number |
| 38 | +of voxels or regions of interest within the brain volume. |
| 39 | + |
| 40 | +In contrast to the symmetric nature of functional connectivity, |
| 41 | +*effective connectivity* denotes asymmetric or causal dependencies |
| 42 | +between brain regions. Here the primary goal is to identify which brain |
| 43 | +structures in a functional network are (causally) influencing other |
| 44 | +elements of the network during some stage or form of information |
| 45 | +processing. Often the term “information flow” is used to indicate |
| 46 | +directionally specific (although not necessarily causal) effective |
| 47 | +connectivity between neuronal structures. Popular effective connectivity |
| 48 | +methods, applied to fMRI and/or electrophysiological (EEG, iEEG, MEG) |
| 49 | +imaging data, include dynamic causal modeling, structural equation |
| 50 | +modeling, transfer entropy, and Granger-causal methods. |
| 51 | + |
| 52 | +Contemporary research on building a human ‘connectome’ (complete map of |
| 53 | +human brain connectivity) has typically focused on structural |
| 54 | +connectivity using MRI and diffusion-weighted imaging (DWI) and/or on |
| 55 | +functional connectivity using fMRI. However, the brain is a highly |
| 56 | +dynamic system, with networks constantly adapting and responding to |
| 57 | +environmental influences so as to best suit the needs of the individual. |
| 58 | +A complete description of the human connectome necessarily requires |
| 59 | +accurate mapping and modeling of transient directed information flow or |
| 60 | +causal dynamics within distributed anatomical networks. Efforts to |
| 61 | +examine transient dynamics of effective connectivity (causality or |
| 62 | +directed information flow) using fMRI are complicated by low temporal |
| 63 | +resolution, assumptions regarding the spatial stationarity of the |
| 64 | +hemodynamic response, and smoothing transforms introduced in standard |
| 65 | +fMRI signal processing (Deshpande et al., 2009a; Deshpande et al., |
| 66 | +2009b). While electro- and magneto-encephalography (EEG/MEG) affords |
| 67 | +high temporal resolution, the traditional approach of estimating |
| 68 | +connectivity between EEG electrode channels (or MEG sensors) suffers |
| 69 | +from a high risk of false positives from volume conduction and non-brain |
| 70 | +artifacts. Furthermore, severe limitations in spatial resolution when |
| 71 | +using surface sensors further limits the physiological interpretability |
| 72 | +of observed connectivity. Although precisely identifying the anatomical |
| 73 | +locations of sources of observed electrical activity (the inverse |
| 74 | +problem) is mathematically ill-posed, recent improvements in source |
| 75 | +separation and localization techniques may allow approximate |
| 76 | +identification of such anatomical coordinates with sufficient accuracy |
| 77 | +to yield anatomical insight invaluable to a wide range of cognitive |
| 78 | +neuroscience and neuroengineering applications (Michel et al., 2004). In |
| 79 | +limited circumstances it is also possible to obtain human |
| 80 | +intracranially-recorded EEG (ICE, ECoG, iEEG) that, although highly |
| 81 | +invasive, affords high spatiotemporal resolution and (often) reduced |
| 82 | +susceptibility to non-brain artifacts. |
| 83 | + |
| 84 | +Once activity in specific brain areas have been identified using source |
| 85 | +separation and localization, it is possible to look for transient |
| 86 | +changes in dependence between these different brain source processes. |
| 87 | +Advanced methods for non-invasively detecting and modeling distributed |
| 88 | +network events contained in high-density EEG data are highly desirable |
| 89 | +for basic and clinical studies of distributed brain activity supporting |
| 90 | +behavior and experience. In recent years, Granger Causality (GC) and its |
| 91 | +extensions have increasingly been used to explore ‘effective’ |
| 92 | +connectivity (directed information flow, or causality) in the brain |
| 93 | +based on analysis of prediction errors of autoregressive models fit to |
| 94 | +channel (or source) waveforms. GC has enjoyed substantial recent success |
| 95 | +in the neuroscience community, with over 1200 citations in the last |
| 96 | +decade (Google Scholar). This is in part due to the relative simplicity |
| 97 | +and interpretability of GC – it is a data-driven approach based on |
| 98 | +linear regressive models requiring only a few basic *a priori* |
| 99 | +assumptions regarding the generating statistics of the data. However, it |
| 100 | +is also a powerful technique for system identification and causal |
| 101 | +analysis. While many landmark studies have applied GC to invasively |
| 102 | +recorded local field potentials and spike trains, a growing number of |
| 103 | +studies have successfully applied GC to non-invasively recorded human |
| 104 | +EEG and MEG data (as reviewed in (Bressler and Seth, 2010)). Application |
| 105 | +of these methods in the EEG source domain is also being seen in an |
| 106 | +increasing number of studies (Hui and Leahy, 2006; Supp et al., 2007; |
| 107 | +Astolfi et al., 2007; Haufe et al., 2010). |
| 108 | + |
| 109 | +In the last decade an increasing number of effective connectivity |
| 110 | +measures, closely related to Granger’s definition of causality, have |
| 111 | +been proposed. Like classic GC, these measures can be derived from |
| 112 | +(multivariate) autoregressive models fit to observed data time-series. |
| 113 | +These measures can describe different aspects of network dynamics and |
| 114 | +thus comprise a complementary set of tools for effective connectivity or |
| 115 | +causal analysis. |
| 116 | + |
| 117 | +Several toolboxes affording various forms of Granger-causal (or related) |
| 118 | +connectivity analysis are currently available in full or beta-release. |
| 119 | +Table 1 provides a list of several of these toolboxes, along with the |
| 120 | +website, release version, and license. Although these toolboxes provide |
| 121 | +a number of well-written and useful functions, most lack integration |
| 122 | +within a more comprehensive framework for EEG signal processing (the |
| 123 | +exceptions being Fieldtrip's routines, and TSA, which integrates into |
| 124 | +the Biosig EEG/MEG processing suite). Furthermore, many of these may |
| 125 | +implement only one or two (often bivariate) connectivity measures, lack |
| 126 | +tools for sophisticated visualization, or lack robust statistics or |
| 127 | +multi-subject (group) analysis. Finally, to our knowledge, with the |
| 128 | +exception of E-Connectome, none of these toolboxes directly support |
| 129 | +analysis and visualization of connectivity in the EEG source domain. |
| 130 | +These are all factors that our Source Information Flow Toolbox (SIFT), |
| 131 | +combined with the EEGLAB software suite, hopes to address. |
| 132 | + |
| 133 | + |
| 134 | + |
| 135 | + |
| 136 | +Table caption. A list of free Matlab-based toolboxes for granger-causal |
| 137 | +connectivity analysis in neural data. |
| 138 | + |
| 139 | + |
| 140 | + |
| 141 | + |
| 142 | +| | | | | |
| 143 | +|------------------------------------------------------------------|------------------|---------|-------------------------------------------------------------| |
| 144 | +| <b>Toolbox Name</b> | <b>Primary Author</b> | <b>Website</b> | <b>License</b> | |
| 145 | +| Granger Causality Connectivity Analysis (GCCA) Toolbox | Anil Seth | <https://www.sussex.ac.uk/research/centres/sussex-centre-for-consciousness-science/resources/connectivity> | GPL 3 | |
| 146 | +| Time-Series Analysis (TSA) Toolbox | Alois Schloegl | <https://sourceforge.net/p/octave/tsa/ci/default/tree/> | GPL 2 | |
| 147 | +| E-Connectome | Bin He | <https://www.nitrc.org/projects/econnectome> | GPL 3 | |
| 148 | +| Fieldtrip | Robert Oosteveld | <http://fieldtrip.fcdonders.nl/> | GPL 2 | |
| 149 | +| Brain-System for Multivariate AutoRegressive Timeseries (BSMART) | Jie Cui | <http://www.brain-smart.org/> | -- | |
| 150 | + |
| 151 | + |
| 152 | +SIFT is an open-source Matlab (The Mathworks, Inc.) toolbox for analysis |
| 153 | +and visualization of multivariate information flow and causality, |
| 154 | +primarily in EEG/iEEG/MEG datasets following source separation and |
| 155 | +localization. The toolbox supports both command-line (scripting) and |
| 156 | +graphical user interface (GUI) interaction and is integrated into the |
| 157 | +widely used open-source EEGLAB software environment for |
| 158 | +electrophysiological data analysis (sccn.ucsd.edu/eeglab). There are |
| 159 | +currently four modules: data preprocessing, model fitting and |
| 160 | +connectivity estimation, statistical analysis, and visualization. First methods implemented include a large number of |
| 161 | +popular frequency-domain granger-causal and coherence measures, obtained |
| 162 | +from adaptive multivariate autoregressive models, surrogate and analytic |
| 163 | +statistics, and a suite of tools for interactive visualization of |
| 164 | +information flow dynamics across time, frequency, and (standard or |
| 165 | +personal MRI co-registered) anatomical source locations. |
| 166 | + |
| 167 | +In this tutorial, we will outline the theory underlying multivariate |
| 168 | +autoregressive modeling and granger-causal analysis. Practical |
| 169 | +considerations, such as data length, parameter selection, and |
| 170 | +non-stationarities are addressed throughout the text and useful tests |
| 171 | +for estimating statistical significance are outlined. This theory |
| 172 | +section is followed by a hands-on walkthrough of the use of the SIFT |
| 173 | +toolbox for analyzing source information flow dynamics in an EEG |
| 174 | +dataset. Here, we will walk through a typical data-processing pipeline |
| 175 | +culminating with the demonstration of some of SIFT’s powerful tools for |
| 176 | +interactive visualization of time- and frequency-dependent directed |
| 177 | +information flow between localized EEG sources in an |
| 178 | +anatomically-coregistered 3D space. Theory boxes throughout the chapter |
| 179 | +will provide additional insight into various aspects of model fitting and |
| 180 | +parameter selection. |
| 181 | + |
0 commit comments