Poster B67, Thursday, August 16, 3:05 – 4:50 pm, Room 2000AB
Evaluating the functional neuroanatomy of language represented across 11,406 neuroimaging studies: a multivariate meta-analysis
Alex Teghipco1, Gregory Hickok1;1University of California, Irvine
Models of the functional neuroanatomy of language are predominantly based on informal evaluations of the literature or formal but domain-specific (e.g., speech production, syntax) meta-analyses. This standard approach fails to consider the relations between functional networks; for example, a "syntax" network might include areas involved in articulatory rehearsal. However, large-scale datasets accumulating neuroimaging findings have made it possible to gauge the selectivity of functional networks by considering how often the same brain areas activate across many different functions (Poldrack, 2011). This development aligns with a growing interest in organizing a cognitive ontology that represents the many-to-many mappings between brain areas and cognitive functions (Price & Friston, 2006; Poldrack & Yarkoni, 2016). To this end, we take a purely data-driven approach to organize and evaluate the language processes embedded in Neurosynth (Yarkoni et al., 2011). We build on prior work describing unique brain networks across task paradigms (Smith et al., 2009; Laird et al., 2011; Yeo et al., 2016) and studies (Poldrack & Yarkoni, 2016), by capturing the core networks underlying a broader class of functional features (i.e. task paradigms, brain areas, methodologies, and cognitive functions attached to studies) while accounting for the selectivity of brain areas that are associated with each feature. Using Independent Principal Component Analysis (IPCA; Yao et al., 2012), we reduce 3,107 meta-analyses of features to 145 networks. This approach generates highly interpretable feature groupings (e.g. memory subsystems, motor control/representation, attentional, emotion/social, and perceptual processes) that are sensitive to subtle dissociations- for instance, phonemic perception and phonological processing dissociate on the basis of involvement in auditory and reading networks. The language-specific networks we uncover include those loading onto lexical-semantics, reading, speech production, auditory perception, and comprehension. With a bootstrap analysis, we show that these networks are amongst the most stable, and can be extracted reliably. Although there is substantial overlap across the language-specific networks, especially in posterior superior temporal sulcus (pSTS) and inferior frontal gyrus, we find strong preferential relations between reading and the fusiform gyrus (FG); lexical-semantics and the inferior temporal gyrus and posterior middle temporal gyrus (pMTG); comprehension and area 55b and pSTS; speech production and ventral premotor cortex as well as pars opercularis; and speech perception and primary/association auditory cortex. We also find that lexical-semantics is considerably more selective for areas of the ventral FG, that reading is selective for anterior portions of the pMTG, and that both auditory and speech production networks are strongly associated with an area in the posterior planum temporale. Further, comprehension exhibits some preference for the anterior temporal lobe, lexical-semantics for anterior pars triangularis and orbitalis, and reading for the paracingulate gyrus and superior parietal lobule. We find that the majority of areas that constitute language-specific networks do not participate considerably in non-language networks. Taken together, these findings highlight a network broadly tuned for language, but comprised of areas that preferentially respond to certain aspects of language function. Moreover, they demonstrate that meta-analytic datasets can be a robust testing ground for theories about the functional organization of language.
Topic Area: Computational Approaches