Modeling suicidal thoughts and behaviors in university students: integrating suicide theory with interpretable machine learning.

Sentence production is the process by which the brain translates from thought to language. Despite its centrality in human cognition, daily life, and in a host of debilitating neurological disorders, it remains poorly characterized in the brain due to a number of critical gaps in the literature. For one, previous research has relied primarily on noninvasive neuroimaging and electrophysiology [e.g., functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and magnetoencephalography], methods which make overt production very difficult to study given their susceptibility to motion and/or motor artifacts. Furthermore, these methods impose a trade-off between spatial and temporal resolution (, ), both of which are needed to study a dynamic, anatomically distributed system such as language. Another (perhaps related) gap is that the neuroscience of language has nearly exclusively focused on either sentence comprehension or single-word production (, ), with only a handful of studies to date investigating the production of multiword utterances (–). One of the complexities of studying sentences is that, relative to single words, they involve higher-order constructs such as combinatorial semantics and syntax. While these systems are rarely investigated with production, they have been extensively studied in the comprehension literature. However, far from elucidating, this body of work contains a number of discrepant findings, and substantial disagreements persist regarding the basic facts. One such area of disagreement is localization: how these systems are spatially encoded. For semantic composition—the process of combining the meanings such as “red” and “boat” into a single representation of a “red boat”—some research points toward a hub in left anterior temporal lobe (–), while other research points toward a broader network spanning left lateral cortex (, , ). For syntax, there is similarly little clarity. Some research has traced syntax back to inferior frontal gyrus (IFG) (, –) with other work identifying posterior temporal areas (, –, –). A recent influential neural model of syntax in the brain () assigns functions to the various regions associated with syntax. This model holds that posterior temporal areas are responsible for forming hierarchically structured representations of a sentence, IFG (particularly pars triangularis) is responsible for transforming these into linear strings of words (or, more specifically, morphemes), and that this is then translated into a sequence of phonemes for articulation by posterior IFG (pars opercularis), dorsal precentral gyrus, and the temporoparietal junction (). However, region of interest (ROI)–based models similar to this face a challenge in accounting for a growing body of findings indicating that, rather than being localized to particular regions, syntax is distributed across broad swaths of cortex (, –). Relatedly, there is substantial disagreement regarding selectivity: the degree to which these systems overlap spatially, which may have important theoretical implications [see (, ) for discussion]. While findings of syntax-selective regions are abundant in the literature (–, , –), recent work using more targeted analytical techniques has reported extensive spatial overlap in these systems (–, –), which has been argued to be evidence that there may, in fact, be no areas of the brain that process syntax to the exclusion of words and/or meaning (). However, other work involving highly sensitive neural measures indicates that spatial overlap and functional selectivity may be able to coexist, enabled by distinct microcircuits in the same region, particularly IFG ().