Meta-analyses

Function/location meta-analysis

Table 18: Mathematical meta-analyses in functional neuroimaging.

Data source	Purpose	Method	Reference
BrainMapTM	Assessment of variation in activation focus	`Functional volumes modeling': Probability density modeling incorporating sample size	[Fox et al., 1997b,Fox et al., 1999,Fox et al., 2001,Fox et al., 1997a]
Lesion database (BRAID)	Determine association between 14 functional variables and 90 brain structural variables in elderly people	Manual delineation of infarct-like lesions. Chi-square contingency table test between pairs of functional and structural variables.	[Letovsky et al., 1998,Herskovits, 2000a]
Lesion database (BRAID)	Determine association between lesion sites and development of secondary attention-deficit hyperactivity disorder (S-ADHD) in children	Manual delineation of lesion from MRI, Mann-Whitney and Fisher exact test statistics in a brain image database (BRAID)	[Herskovits et al., 1999,Herskovits, 2000a]
Lesion database (BRAID)			[Herskovits, 2000b]
Literature	Determination of areas involved in language production		[Indefrey and Levelt, 2000,Indefrey and Levelt, 1999]
Literature (25 studies)	Distinction between dorsolateral and frontopolar cortex	Hotelling's $T^2$ statistics used to compare two groups of activations in Talairach space. $\chi^2$ used for region test.	[Christoff and Grabrieli, 2000, page 176]
Literature	Determining of different activation between 5 sets of activation foci each set involving different cognitive demand	Multidimensional Kolmogorov-Smirnov applied multiple times.	[Duncan and Owen, 2000]
Literature (28 studies)	Determination of areas involved in syntactic parsing	Region-based analysis (102 areas) with $P$-value threshold determined from a binomial distribution.	[Indefrey, 2001]
BrainMapTM	Novelty detection in nomenclature	Probability density modeling through adaptive kernel density modeling conditioned on anatomical labels.	[Nielsen and Hansen, 2002c,Nielsen et al., 2001]
Literature (9 studies)	Prediction of areas involved in single word reading	Probability density modeling through kernel density modeling of 154 Talairach coordinates	[Turkeltaub et al., 2002,Turkeltaub et al., 2001]

Table 18 displays some of the meta-analyses that use mathematics/statistics. For an early review see [Fox et al., 1998]. For a more recent overview see [Wager et al., 2007].

Meta-analysis of Talairach coordinates was pioneered by Peter T. Fox et al. under the name ``functional volumes modeling'' (FVM) [Fox et al., 1997b,Fox et al., 1999,Fox et al., 1997a]. These original studies used parametric Gaussian models. Non-parametric modeling of the distribution of brain foci was first described in two unpublished studies with Gaussian mixture modeling [Nielsen and Hansen, 1999] and adaptive Gaussian kernel density modeling [Nielsen and Hansen, 2000]. Later studies include adaptive Gaussian kernel density modeling for database outlier detection [Nielsen and Hansen, 2002c,Nielsen et al., 2001,Nielsen et al., 2000], Gaussian kernel density modeling in connection with single word processing [Turkeltaub et al., 2002,Turkeltaub et al., 2001], kernel density estimation in connection with Broca's area and verbal working memory [Chein et al., 2002,Chein et al., 2001], kernel density modeling with a spheric uniform kernel in emotion [Wager et al., 2003], and a model combining kerndel density modeling with a Gaussian mixture model [Neumann et al., 2008]. Another study is [Wager et al., 2004]. A large number of similar studies appears in a special issue of the Human Brain Mapping, volume 25, issue 1 [Fox et al., 2005b]. Functional volumes modeling is sometimes referred to as ``voxelization'' or ``activation likelihood estimation'' (ALE). Some form of meta-analysis with the use the BrainMap^TM database has been briefly described [Mahurin et al., 1995].

Determining statistical thresholds in one set of voxelized Talairach coordinates is described in [Turkeltaub et al., 2002,Turkeltaub et al., 2001] and [Nielsen, 2005]. Statistical methods for determining whether two sets are different are described in [Christoff and Grabrieli, 2000,Duncan and Owen, 2000] and [Nielsen and Hansen, 2004a,Nielsen et al., 2004b,Nielsen et al., 2005,Nielsen et al., 2004a] and [Laird et al., 2005a]. The multidimensional Kolmogorov-Smirnov used in [Duncan and Owen, 2000] is originally from [Peacock, 1983] and is also described and implemented in [Press et al., 1992, section 14.7]. Hotelling's $T^2$ test was also used in [Berman et al., 1999, page 212] but not in connection with a meta-analysis. Statistical tests on warped coordinates are described in [Steel and Lawrie, 2004].

The Brede Toolbox automatically performs multivariate analyses such as singular value decomposition (principal component analyses), independent component analyses, non-negative matrix factorization and K-means on voxelized Talairach coordinates on the entire Brede Database [Nielsen, 2003,Nielsen et al., 2004c]. Another multivariate analysis method, ``replicator dynamics'', is suggested in [Neumann et al., 2005].

A number of meta-analytic studies have grown out of the BRAID database: [Herskovits et al., 1999,Megalooikonomou et al., 1999,Megalooikonomou et al., 2000,Megalooikonomou and Herskovits, 2001,Lazarevic et al., 2001,Herskovits et al., 2002].

Descriptive statistics of activation foci appears in [Markowitsch and Tulving, 1994], where the fraction of fundus activations over 30 PET studies is found.

Function/location meta-analysis without spatial information other than the text can also extract relevation associations. Name entities has been extracted from well over 100,000 PubMed abstracts with the use of Unified Medical Language System (UMLS) and manually developed rules for rule-based name entity extraction [Hsiao et al., 2007]. This system extracted phrases/terms for brain function and neuroanatomy and built an interactive visualization system to display the function-anatomy graph.

Table 19: Meta-analysis in Talairach space of brain function. KDE is kernel density estimation (in Talairach space).

.
Area	Function	Method	Description	Reference
Left inferior frontal cortex	Semantic and phonological processing	Tables, plots	Phonological processing dorsally while semantic ventrally	[Poldrack et al., 1999]
Anterior cingulate	Cognition, emotion			[Bush et al., 2000]
Many	Cognition	Tables, plots		[Cabeza and Nyberg, 2000]
Inferior frontal	Phonological processing	Plots		[Burton, 2001]
Prefrontal	Cognition, emotion	Plots, warp transformation, MANOVA, KDE	Resampling was used for significance test	[Steel and Lawrie, 2004]
Orbitofrontal		Plots		[Kringelbach and Rolls, 2004]
Medial frontal	Self/Other	Clustering(?)		[Seitz et al., 2005]
Posterior cingulate	As many as possible	Text clustering, Hotelling's test	Text mining on PubMed abstract for clustering articles. Thereafter determination of segregation between coordinates in clustered articles	[Nielsen et al., 2005]

Connectivity analyses

Table 20: Connectivity analyses

Species	System	Reports	Areas	Conn.	Levels	Method	References
Rat	Hippocampus	$>900$ (14000)	23		0-3, c, x	2D and 5D MDS (from SAS MDS), Cluster analysis (SAS 6.09 MODECLUS), Venn diagram	[Burns and Young, 2000]
Macaque	Cortical sensory	[Felleman and Van Essen, 1991]	30/14		$<, \leq, \emptyset, \geq, >$	Hierarchical analysis	[Hilgetag et al., 2000b]
Cat	Cortical sensory	[Scannell et al., 1995]	22		$<, \leq, \emptyset, \geq, >$	Hierarchical analysis	as above
Cat	Cortical	Part of [Scannell et al., 1999]	55	892	0-3	Optimal set analysis, MDS (SAS MDS), Cluster analysis (SAS MODECLUS), small-world coefficient	[Hilgetag et al., 2000a]
Macaque	Visual	[Felleman and Van Essen, 1991]	32	319		as above	as above
`Primate'		[Young, 1993]	73	834		as above	as above
Macaque	Somatosensory motor	[Felleman and Van Essen, 1991]	15	66		as above	as above
Primate	Cortical	19 (CoCoMac-Stry)	39	``3897 tests''	0-3	Optimal set analysis, MDS (SAS MDS), small-world coefficient	[Stephan et al., 2000a]

An early program for connectivity analysis is ``Connection'' [Nicolelis et al., 1990].

[Kaiser and Hilgetag, 2004] used data from CoCoMac together with spatial positions from Caret to get an approximation for the wiring length in cortex.

[Toro and Paus, 2006] described a co-activation analysis of functional activation recorded in the BrainMap^TM database and constructed a program for interactive visualization of the 6-D probabilistic map. A smaller co-activation study analysing data from 126 papers focused on connections from the basal ganglia [Postuma and Dagher, 2006].

A database with volumes for anatomo-functional connectivity might become available [Poupon et al., 2006].

There is a number of studies using the connectivity of the small worm Caenorhabditis elegans (WormAtlas), e.g., for examining the small-world phenomenon [Watts and Strogatz, 1998], or explaining the neuronal placement [Chen et al., 2006,Ahn et al., 2006] and (Kaiser and Hilgetag, 2005, Society for Neuroscience).

Unclassified

Meta-analysis of ERP in schizophrenia [Bramon et al., 2001].

[Young and Scannell, 2000]