Executive functions constitute higher-level cognitive processes that influence lower-level automatic processes to cope with non-routine situations1 and improve cognitive efficiency in routine situations.2 Despite their critical contribution to adaptive processes and human activities, the functional organization and related anatomy of executive functions are still largely unknown. The main control processes have been defined, along with their critical tasks and performance patterns.2,3 Conversely, actions exerted by control processes on lower-level automatic processes are still poorly defined. However, they are necessary in the design of the overall cognitive architecture (i.e., the pathways followed by information from entry to output). Consequently, the contribution of each cerebral region found to be associated with executive tasks is still only poorly defined, leading to an imprecise understanding of the anatomy of executive functions.

We probed executive functioning by examining the cognitive architecture and voxelwise anatomy of the verbal fluency task, i.e., a speed- and strategy-based lexico-semantic oral production task. This constitutes a leading executive task4, sensitive to major brain diseases.2

This task involves one key executive process, specifically the generation of non-overlearned responses. However the processes involved in this task and more important, their links with other functions, such as language (i.e., the functional architecture) remain poorly and indirectly addressed. The contribution of several processes has been proposed such as (1) search by clustering (i.e., identifying subsets within the general category, e.g., bird names for animal semantic fluency) and switching from one cluster to another,5 (2) working memory which is required to temporarily retain the produced words in order to avoid their repetition,6,7 (3) inhibition which would be needed to avoid rule-breaking errors,6,8 (4) attention and processing speed that is needed to produce as many as possible words in this time-constrained task.8 Despite their importance these results remain subject to insufficient control of confounding factors, especially due to assessment of working memory, inhibition or processing speed using complex tasks or tasks involving linguistic component. Moreover association analyses usually do not control for the effect of demographic factors which artificially increases the strength of association.9 Finally none of these studies has proposed a functional architecture of executive processes and their relationships with other functions that might account for verbal fluency task. More specifically the interaction between strategic lexico-semantic search and semantic and lexico-phonological processes is still undefined.

The voxelwise anatomy of the fluency task has been examined using functional activation MRI (fMRI)10 with meta-analysis11 and, to a lesser extent, voxelwise lesion-symptom mapping (VLSM) in stroke patients.8,12, 13, 14, 15, 16, 17 These studies reported the role of left lateral frontal gyri, dorsal cingulate, inferior parietal lobule, middle and inferior temporal gyri, insula, and deep grey and white matter (WM) structures.12, 13, 14, 15, 16, 17 Attempts to separate the anatomy of semantic and letter fluency have provided conflicting results, with a prominent role of frontal structures in letter fluency and a prominent temporal role in semantic fluency reported in some studies12,15,16 but not others.14,17 As most of these regions are also associated with oral language output, the structures that account for strategic search process are still largely unknown.

To address these limitations, this study proposed and then evaluated a modeling of the verbal fluency task and finally examined its anatomy. According to the model of Ellis and Young18, linguistic processes involved in oral language production include semantic memory, an output lexicon, a phonological-assembly buffer, and the articulation of language (Figure 1). The confrontation naming task involves this oral language production pathway.18, 19, 20 We hypothesized that the fluency task additionally triggers two control processes required for a rapid strategic search (Figure 1): (1) a specific strategic search process selecting and activating words from a semantic memory-lexicon for semantic fluency and from an output lexicon for letter fluency and (2) a general attentional activation process required to accelerate the processing speed.21 According to this model, fluency performance is related to the interaction of three sets of processes: (1) lexico-semantic storage and oral language output processes, which are also involved in confrontation naming (indexed by naming performance), (2) a strategic search process, that is not directly assessable, and (3) an attentional activation system that optimizes processing speed. Processing speed was specifically examined with part A of the Trail Making test, a simple visuomotor speed test devoid of language and complex processing.2 This model predicts three findings: (1) a relationship between fluency and both naming and processing speed; (2) if the strategic search process accounts for fluency performance, then residuals of fluency performance after partitioning out variance related to naming and processing speed (that are considered to estimate performance of search process) may share common variance across fluency tests and may be represented by a latent factor in a structural equation modeling (SEM) analysis; and (3) if strategic search process accounts for fluency impairment of patients, then their residuals of fluency performance should be impaired. In addition if such strategic search process depends on dedicated brain area, then contrast between tasks (fluency, naming and processing speed) should identified an additional area specific to fluency task.

We sought to determine the cognitive architecture and voxelwise anatomy of the fluency task using the large dataset of the GRECogVASC cohort, which includes stroke patients22 and healthy controls23 assessed with a standardized battery whose performance scores were analyzed using a validated method.9 Voxelwise anatomy was determined for patients using a previously validated multivariate approach based on VLSM24,25 and disconnectome analyses.26 The results were crossed with fMRI-based meta-analytic data.27