Dorsal Medial Prefrontal Cortex Plays a Necessary Role in Rapid Error Prediction in Humans

Discussion

An extensive functional imaging literature argues for a central role for dACC in cognitive control. This role has been variously cast as monitoring response conflict (Botvinick et al., 2001), acting as a response comparator (Ullsperger et al., 2005), detecting errors (Holroyd et al., 2005), predicting error likelihood (Brown and Braver, 2005), engaging attention to overcome prepotent response tendencies (Pardo et al., 1990; Posner and DiGirolamo, 1998; Peterson et al., 1999), or representing action-outcome contingencies (Rushworth et al., 2004; Williams et al., 2004; Rushworth et al., 2007). Several of these accounts have been formalized in computational models (Botvinick et al., 2001; Brown and Braver, 2005, 2008; Holroyd et al., 2005). In this increasingly tightly specified area of investigation, built on extensive evidence from both functional imaging and electrophysiology experiments, the absence of unequivocal loss-of-function evidence in support of any of these hypotheses has been troubling. The present study focused on error-related processes, and provides evidence that dorsal mPFC is necessary for rapid error prediction.

These findings argue that there are multiple routes to recognizing that an error has occurred. Dorsal mPFC appears to be critical only for the rapid, on-line identification of an error, before (or during) its occurrence. Reporting of an error after it had occurred, over a time course that is consistent with registering the error after the motor response had been produced, remained intact despite dorsal mPFC damage. Early behavioral work established that error correction in choice reaction time tasks occurs more rapidly than the choice reaction time itself, and in some paradigms, is also more rapid than the simple reaction time (Rabbitt, 1966; Angel, 1976). This form of rapid error correction is therefore thought to rely on monitoring of an efference copy of the on-going response. The most intensively studied error-related ERP component, the ERN, peaks within 100 ms of the motor response, and likely arises from dorsal ACC (Dehaene et al., 1994; Herrmann et al., 2004; Vocat et al., 2008). The time course of the ERN also suggests that it relies on monitoring of an efference copy of the motor plan, rather than sensory feedback (Gehring et al., 1993). The present findings support the existence of a distinct error monitoring process relying on efference copy, and provide strong evidence that a region within dorsal mPFC, most likely dorsal ACC, plays a necessary role in this process, either at the point of error detection, or in permitting a rapid corrective response.

The findings from the modified n-back task further suggest that this region is necessary for the conscious evaluation of error likelihood. Thus, mPFC seems to play a critical role in the within-trial modulation of a response by the likelihood that response is an error, a process that has an effect on the readiness to produce the alternate response, and a “conscious” correlate reflected in the overt confidence in the response. These findings together are strongly supported by a functional magnetic resonance imaging (fMRI) study which aimed to disentangle conscious, pre-error performance monitoring from the processes involved in error commission (Magno et al., 2006). The task used in that study allowed subjects to avoid errors by rejecting trials they thought they might fail. Dorsal ACC activity was greater for such “reject” trials, compared with trials in which an error was made. This was interpreted as supporting a role for dACC in “on-line adjustment of behavior to prevent errors” (Mango et al., 2006). Our results indicate that dorsal mPFC plays a necessary role in such a process.

The data presented here do not speak to whether the pre-response processes mediated by dorsal mPFC are engaged only during error trials, or are present during trials with response competition more generally, although both ERP and fMRI studies suggest the latter (Carter et al., 1998; Botvinick et al., 1999; Van Veen and Carter, 2002). The tendency for those with mPFC damage to have larger flanker effects could be interpreted as reflecting difficulty resolving response conflict even on correct trials. However, in this particular study, this effect should be interpreted with caution, given that task parameters were not identical across patients, or between the patient and control groups. Further, it is striking that although all five patients showed slowing of error correction, the size of the flanker effect was more variable, falling well within the normal range in two of five patients.

In general agreement with the current findings, a previous case report of a patient with a very circumscribed lesion to dACC reported a small, but significant decrement in the rate of error correction in a Stroop variant, with intact post-error slowing (Swick and Turken, 2002). Reaction times for error corrections were not reported, however.

The single dACC-damaged patient reported by Swick and Turken had intact conflict monitoring effects in the Stroop task, despite reduced error correction, further suggesting that error and conflict monitoring do not share a reliance on intact mPFC (Swick and Turken, 2002). At the group level, the current findings, in conjunction with a prior study showing intact conflict monitoring in the Stroop task in a partly overlapping group of patients (Fellows and Farah, 2005) (three patients participated in both studies), also argue for a distinction between error prediction and conflict monitoring, with dorsal mPFC necessary only for the former. However, it may be that the usual behavioral measures of conflict monitoring are more sensitive to post-conflict adjustments (which the present data suggest are likely to be intact despite mPFC damage) rather than on the within-trial anticipation of conflict that would be the direct homolog of the error effects we report here.

Interestingly, a recent group study found that sequential (i.e., trial-by-trial) congruency effects in the Simon task were not reliably present after ACC damage (di Pellegrino et al., 2007). These patients also failed to show consistent post-error slowing. In contrast to flanker and Stroop tasks, where cognitive control is likely exerted via selective attention to the relevant stimulus feature (and in turn its associated response), the Simon task more directly taps conflict at the response level (Liu et al., 2004). This distinction, perhaps in addition to differences in precise lesion location (the explanation advanced by di Pellegrino et al.), may be important in reconciling these conflicting findings.

Although several of the patients we studied here had quite extensive mPFC (and in two cases, orbitofrontal cortex) damage, the common area of overlap encompasses dACC, making this region the most likely driver of the slowed error correction that was present in all five patients. Nonetheless, we cannot entirely exclude the possibility that direct damage to, or disruption of connections with nearby regions, including rostral ACC and pre-supplementary motor areas, or indeed disruption of longer range connections, such as via the cingulum bundle or the genu of the corpus callosum, contributes to the effect.

Other recent work has argued for a more nuanced view of dACC's role in response monitoring, one that can account for a putative role for this region in action-outcome learning and related value-related decisions. Our findings indicate that this region is critical for the prediction of response errors, specifically. Further work will be required to determine whether this generalizes to a necessary role in the prediction of the relative value of actions more generally, in contexts where responses have a graded value, rather than being dichotomously “right” or “wrong.” Other lines of evidence support this possibility (Gehring and Willoughby, 2002; Rushworth et al., 2004; Quilodran et al., 2008), and one study of the effects of cingulotomy found an increase in errors when response selection was guided by a reduction in reward (Williams et al., 2004). Furthermore, we cannot speak to whether this region is necessary for learning from errors, except to note that one potential index of trial-to-trial adjustment, post-error slowing, was intact despite mPFC damage.

This demonstration of a critical role for dorsal mPFC in only one of several measured error-related processes highlights an important point. Error-related processes are obviously critically important to adaptive behavior, and it seems highly likely that there are multiple neural systems that encode error-related information. Such duplication might compensate for lesion effects, as was demonstrated by the apparent use of “error detection” mechanisms to both report and correct errors after mPFC damage in patients studied here. Deployment of particular error and conflict-related processes appears to differ across individuals, and across task demands (Garavan et al., 2002; Hester et al., 2005; Brown and Braver, 2007). Behavioral measures that focus on the timing of these processes may be helpful in understanding these different mechanisms.

Just as there are likely multiple neural systems involved in error processing, a given neural circuit might be engaged by error prediction in a specific response domain, or serve a more general performance monitoring role across different response types. We recently evaluated memory monitoring in the same five patients studied here in a separate set of experiments. Interestingly, they showed clear deficits in only the prospective, within-trial ability to predict episodic memory performance, as captured by the accuracy of a “feeling-of-knowing” measure (Modirrousta and Fellows, 2008). This supports a more general error likelihood account of dorsal mPFC function, consistent with fMRI findings (De Martino et al., 2006; Magno et al., 2006; Paulus and Frank, 2006; Brown and Braver, 2007), but also helps to pinpoint the timing of the critical contribution of this region to such monitoring as within-trial and before the response.

In summary, damage involving dorsal mPFC in humans selectively disrupted anticipatory error monitoring, reflected in the speed of error correction and the conscious prediction of errors, while sparing post-response error reporting, and performance adjustment on post-error trials. This result provides important converging evidence for error prediction accounts of the function of this brain region.