Lesion mapping of cognitive control and value-based decision making in the prefrontal cortex

Results

We selected five neuropsychological target scores: four that emphasize cognitive control and one that measures value-based decision-making.

The four cognitive control tasks were as follows: the Part B – Part A difference score from the Trail-Making Test (TMT), a measure of executive response switching; the Perseverative Errors score from the Wisconsin Card Sorting Test (WCST), which measures impairments in set switching; the Color-Word Interference score from the Stroop Test (STROOP), a measure of response inhibition; and the Number of Words score from the Controlled Oral Word Association Test (COWA), which measures verbal fluency, divergent thinking, and response creativity. As an index of value-based decision-making and reward learning, we used the Net Score (advantageous minus disadvantageous choices) from the Iowa Gambling Task (IGT) (see SI Materials and Methods for a description of the neuropsychological tests). All these tasks have been extensively used and well standardized, and they have been shown to detect impairments reliably in clinical populations such as ours. As expected, the cognitive control-related tasks were all weakly, but positively intercorrelated (r = 0.16–0.43), whereas their correlation with the IGT was generally lower (r = 0.09–0.20) (Table S1), a pattern that remained even after the covariates were statistically removed from the data (see below).

To isolate cognitive control and value-based decision-making, we removed the covariance of four basic cognitive skills that our tasks also draw on (2) (verbal abilities, spatial abilities, verbal memory, and spatial memory) by using multiple linear regression (Materials and Methods and SI Materials and Methods). All subsequent analyses were based on these covariate-corrected task residuals, which were then submitted to lesion overlap analyses by using nonparametric VLSM (28). This analysis compares the scores at every voxel in the brain (corrected for multiple comparisons) between that subset of patients whose lesion includes a given voxel versus that subset whose lesion does not. Lesion density maps (Fig. S1) and power analyses (Fig. S2) confirmed that we had sufficient lesion coverage in the PFC, and most other regions of the brain, to detect significant lesion-deficit relationships on our target tasks (see Fig. S3 for further quantification of lesions in the PFC).

Performance on tasks engaging cognitive control was associated with dorsal sectors of the medial PFC and both ventral and dorsal sectors of the lateral PFC (Fig. 1A). Significantly lower performances on the TMT and WCST were associated with damage to the ACC. Lower performance on the TMT was associated with damage to a focal region of left rostral ACC, whereas lower performance on the WCST was associated with damage to a slightly more anterior region of the ACC and left medial superior frontal gyrus, together with multiple right hemisphere regions including temporal (superior temporal gyrus, posterior middle temporal gyrus), parietal (angular gyrus), and frontal (precentral gyrus) cortices, and subcortical regions (head of the caudate). The commonalities across these two tasks suggest that the rostral ACC is critically involved in response/set switching (29), whereas related executive functions such as error (30) and conflict detection (27, 31) have been associated with the (more posterior) dorsal ACC. Lower performance on the STROOP was associated with damage to left dlPFC (middle frontal gyrus), which is consistent with the previously described role of this region in response inhibition (23, 30, 32). Lower performance on COWA was associated with damage to extensive sectors of the left frontoparietal cortices, anterior PFC, and insula, an intriguing finding considering that we had already removed verbal ability and memory from task scores (33). The pattern of relative lateralization across cognitive control tasks is consistent with right hemisphere involvement in error monitoring, particularly with conceptual responses (e.g., WCST), and left hemisphere involvement in setting and maintaining task context (e.g., the different instructional conditions of the Stroop) (6, 34).

In contrast with the above findings, the Iowa Gambling Task (IGT) was the only task associated with ventral sectors of the medial PFC, lateralized to the left. The IGT was also associated with dorsal sectors of the anterior PFC on the right, and with the ACC, frontal pole, and the superior and middle frontal gyri both medially and laterally (Fig. 1B). Our vmPFC finding is consistent with previous lesion work in small samples (e.g., refs. 16 and 35), including studies using other gambling tasks that separate risk-taking from motor impulsivity (36, 37), and with functional imaging studies on reward learning (18, 38–40). However, the recruitment of additional PFC regions varies to some extent depending on the particular gambling task administered (37, 41). We have noted (10, 22) that selective impairments on the IGT can result from damage to the vmPFC, and we (34) and others (42) have also pointed out that damage to other sectors of the PFC involved in working memory can result in impaired IGT performance.

To visualize the overlap and uniqueness of lesion-deficit effects across all tasks, we projected our primary statistical VLSM results from the large sample [P < 0.05, false discovery rate (FDR)] onto a single template brain (Fig. 2A) and computed the extent of their pairwise anatomical overlap (Fig. 2B). Most tasks showed remarkable anatomical specificity, with near-zero overlap ratios, with two notable exceptions: TMT and WCST shared a common locus in the rostral ACC, and the significant effect for STROOP in the dlPFC was completely included in the large lesion effect for COWA. Interestingly, IGT showed essentially zero overlap with any of the cognitive control tasks (Fig. 2B), consistent with the conclusion that brain systems implicated in cognitive control versus value-based decision-making in the PFC are at least partially dissociable anatomically. The findings demonstrate that largely nonoverlapping sectors of the PFC subserve cognitive control and valuation, respectively.

The sample sizes for these findings varied to some extent across different tasks (Table S1, bold type) because of inclusion of all participants to maximize statistical power and brain coverage. We next analyzed results from a subset of 62 patients who had complete data on all five target tasks. This analysis is important because it removes the possibility that differing sample sizes might contribute to the dissociations we reported. We carried out the same VLSM analyses in this smaller sample within the same regions found earlier (Fig. S4). Relative to the task-related regions identified in the larger sample, this analysis corroborated the primary findings for the TMT, IGT and, to a lesser degree for WCST, STROOP, and COWA, because of the more exclusive focus of lesion densities in the anterior PFC.

The above findings show that specific sectors of the PFC are required for the performance of different executive functions. Because we used multiple tasks to assess cognitive control (TMT, WCST, STROOP, COWA), we further investigated whether there might also be brain regions shared among all tasks measuring cognitive control. Using data from 104 patients who had complete datasets on all four cognitive control tasks, we conducted a factor analysis on the same residualized scores that were entered into the previous analysis and analyzed the resultant single “cognitive control factor” (factor loadings: TMT = 0.60, STROOP = 0.52, COWA = 0.49, WCST = 0.43) with VLSM (Fig. 3A). Significantly lower performance on this factor was associated with damage to the left rostral ACC, at a location partially overlapping with the location found to be associated with lower scores on the TMT and WCST (Fig. 3B). Other significant effects associated with this cognitive control factor were observed in the left precentral gyrus (ventral to the location of effects for STROOP), middle OFC and putamen, and the right globus pallidus. Thus, despite clear differences in the association of lower performance in the individual tasks with location of damage, the results of this analysis suggest that these tasks also all draw upon a cognitive control component that critically depends on the rostral ACC.

To obtain a summary view of the distinct prefrontal networks for cognitive control and valuation, we returned to our subsample of 62 patients who had scores on all tasks and projected their mean within-subject difference scores on the IGT and the common cognitive control factor onto a template brain (Fig. 3C). This analysis delineated a medial network consisting of vmPFC, frontal pole, and medial frontal cortex associated with valuation (i.e., patients with lower IGT scores) (14, 17) and a right-lateralized lateral network including ventro- and dorsolateral PFC associated with cognitive control (i.e., patients with lower cognitive control factor scores) (4, 11, 32, 43–45). Interestingly, the rostral ACC exhibited difference values around zero, which suggests that this region is not clearly dominated by either cognitive control or valuation.

Discussion

This comprehensive assessment of lesion-related deficits in cognitive control and value-based decision-making demonstrates a robust anatomical specificity of cognitive functions within the PFC made possible through our assessment of performance across multiple tasks. Our findings support the conclusion that the rostral ACC plays an essential role in flexibly shifting between cognitive tasks and response sets, whereas lateral structures in the dlPFC play an essential role when competing responses need to be inhibited (27, 44). These regions contribute to a control network (Fig. 3C, red areas) that maintains goals by flexibly adjusting attentional and working memory resources to changing environments and task demands (4). In contrast, the left vmPFC is a critical component of a valuation and reward learning network (Fig. 3C, blue areas) that is concerned with evaluating incoming stimuli and computing their expected future reward to guide choices (17). These findings provide comprehensive lesion-based evidence for a high degree of functional-anatomical specificity in the human PFC.

Our findings also provide important insights into the component processes of cognitive control. Within the left ACC, only TMT and WCST shared directly overlapping neural sectors, consistent with the correlated demands made by these two tasks in flexibly switching between response or instruction sets (45). In fMRI studies, activity in this shared region of the rostral ACC has been related to uninstructed set-shifting as it occurs in the WCST (46) and to error detection (30), although WCST-related activations also often occur more dorsally and in conjunction with a frontoparietal network involving executive attention and working memory (44). In the macaque, lesions to the anterior sectors of the cingulate sulcus impair rule switching (47). Interestingly, a recent human neuroimaging study showed that activity in the rostral ACC is modulated by estimates of the volatility of outcome contingencies (48), a process akin to tracking (and shifting) between tasks or stimulus sets, as tested in the WCST and TMT. Computationally, the ACC is thought to accomplish these functions with the aid of striatal dopaminergic input that facilitates the representation of action values and prediction errors in the ACC (49), both of which are important for detecting shifts in environmental contingencies and adapting ongoing behavior accordingly. The division of labor between these two networks has its neurobiological correlate in the ventromedial and dorsolateral sectors of the PFC, highlighted in the difference image shown in Fig. 3C. The interaction between these two networks converges in the ACC, which exhibits difference values around zero, suggesting that it is dominated neither by the control nor the valuation network. Our overlap analysis of the cognitive control factor (Fig. 3 A and B) suggests that the key function of the rostral ACC may be set shifting, whereas more posterior subregions within the dorsal ACC, as well as areas of right dlPFC, may be recruited for functions such as error detection (30) and conflict monitoring (27, 31) that are important for cognitive flexibility (6).

Lower performance in the STROOP was only associated with lesions in the left dlPFC, a finding consistent with previous lesion studies (23, 25), but only partly consistent with early neuroimaging studies (e.g., ref. 27), which emphasized response conflict detection in the dorsal ACC as the critical component for Stroop performance and assigned a role of inhibitory control to the (left) dlPFC. The consistent implication by neuroimaging studies of these two regions in Stroop performance has been supported by a meta-analysis (32). A possible explanation that reconciles these findings is that the behavioral reaction time differences, which constitute the principal Stroop summary score, capture the increased effort of inhibiting prepotent responses, whereas fMRI activation at the time of stimulus presentation reflects the response conflict induced by interference items. Another possibility is that the ACC is only recruited for Stroop performance under the unblocked stimulus presentation that is typical of fMRI studies, contrary to most lesion studies that use the standardized blocked presentation of the Stroop items (26).

There was a strong effect of left vmPFC damage on performance in the Iowa Gambling Task. IGT scores are calculated as the difference between the number of choices from two advantageous (net positive reward outcome) minus two disadvantageous (net negative reward outcome) decks, a measure of how well participants learn the expected value of the decks to guide their choice. Furthermore, impaired performance on the IGT and the extent to which lesions overlapped with the vmPFC were significantly correlated (r = −0.37, P = 0.0001) (Fig. 2C). Our findings are consistent both with previous lesion studies in smaller samples (e.g., ref. 35), and with several fMRI studies (14) that suggest that the vmPFC represents the expected reward value of a choice (38, 39). Similar conclusions are supported by electrophysiology (17) and lesion studies (50) in the monkey. Our findings are also consistent with lesion findings reporting that damage to the vmPFC can spare performance on the WCST (51) and TMT (24).

The IGT is a widely used valuation task (21), but to broaden the sampling of behavior in the domain of value-based decision-making, we identified a subset of our patients who have been tested on other value-based decision-making tasks, albeit in different studies (52, 53). These patients, whose lesions encompassed our left vmPFC “hotspot” for the IGT (Fig. 1B), exhibited specific decision-making deficits in the Ultimatum Game (52) and the Explore-Exploit task (53) and were also impaired on the IGT (Fig. 2C), supporting a general role of this area in value-based decision-making that generalizes beyond the specific demands of the IGT.

The connectivity of the vmPFC (as distinct from that of the dlPFC; ref. 7) positions it well to integrate and represent the value that is expected from a choice (54). Interestingly, the IGT also showed an effect in the right fronto-polar cortex extending into the right ACC and the middle and superior frontal gyri. The former is commonly associated with highly abstract planning and subgoal processing and is part of a rostro-caudal hierarchy of abstraction (11). There is also evidence for a distinction in frontopolar cortex between more medial sectors that subserve stimulus-oriented attention (SOA) versus more lateral sectors that subserve stimulus-independent attention (SIA) (50). The lesion effect for the IGT in our study encompasses both of these sectors, perhaps indicating that the task requires both attention to the card decks (SOA) and the updating of each deck’s values (SIA).

The ACC has an established role in error detection and conflict monitoring (27, 30, 31). These regions may reflect interactions between the valuation network and the cognitive control network in performance of the IGT. The predominantly left-lateralized vmPFC effect we found may result from interactions with sex: Whereas both men and women showed strong lesion effects for the IGT in the left vmPFC, only men also showed an effect in the right vmPFC (Fig. S5), consistent with prior studies (42, 55). Alternatively, the left-lateralized IGT effect may be due to effector-specific, contralateral value representations in the vmPFC in our mostly right-handed participants (317 of 344) (56).

Our study involved a large sample of patients, but even so, it is important to qualify the findings by considering the heterogeneity of the distribution of lesions. Although we had sufficient statistical power to detect lesion-deficit associations over most of the brain in principle (Fig. S2), power was greatest in those regions with the densest lesion overlaps for a task. In all cases, this region was the prefrontal cortex, making our conclusions particularly robust with respect to the sectors of PFC that we describe. However, our findings do not rule out (and in some cases support) a role for structures outside the frontal lobes in the cognitive tasks we studied. Thus, although we demonstrate an anatomical dependency on and dissociation of functions for particular sectors of the PFC, there is little question that these regions implement cognitive control and decision-making as part of larger neuroanatomical networks.

We would hasten to add that the structure-function mappings we report here by no means exhaust the functions of the PFC, which is known to participate in other high-level cognitive processes such as theory of mind (1) and self-referential processing (57, 58). Also, it is important to acknowledge that our results are relative to the particular tasks for which data were available, and different neuropsychological measures might produce somewhat different anatomical results. However, we used clinically relevant tests with well-established construct validity (21), and provided that underlying constructs were kept constant, we would not expect findings from other tests to differ in major ways from those that we obtained. Moreover, the definitiveness of the brain–behavior relationships we report here is enhanced by our uniquely large sample size.

Finally, important open questions remain regarding how connectivity between ventromedial and dorsolateral regions of the PFC implements their network functions, how the roles of these regions and networks may differ across individuals and contexts, and how deficits arising from lesions within them may be partly compensated by plasticity and reorganization. The present set of findings provides a comprehensive description of the core sectors of the PFC, on the foundation of which such questions could be investigated in future studies.

Materials and Methods

For brevity, we provide an overview of the materials and methods. A detailed description can be found in SI Materials and Methods.

Supplementary Material