Time-sensitive prefrontal involvement in associating confidence with task performance illustrates metacognitive introspection in monkeys

Abstract

Metacognition refers to the ability to be aware of one's own cognition. Ample evidence indicates that metacognition in the human primate is highly dissociable from cognition, specialized across domains, and subserved by distinct neural substrates. However, these aspects remain relatively understudied in macaque monkeys. In the present study, we investigated the functionality of macaque metacognition by combining a confidence proxy, hierarchical Bayesian meta-d' computational modelling, and a single-pulse transcranial magnetic stimulation technique. We found that Brodmann area 46d (BA46d) played a critical role in supporting metacognition independent of task performance; we also found that the critical role of this region in meta-calculation was time-sensitive. Additionally, we report that macaque metacognition is highly domain-specific with respect to memory and perception decisions. These findings carry implications for our understanding of metacognitive introspection within the primate lineage.

Fig. 1. Task performance and metacognitive capability remained steady across days.

Plots depict daily accuracy (a, c) and metacognitive efficiency (b, d) across 20 days for four monkeys performing two tasks. Strong correlations between the two meta-cognitive metrics (e, f). Pearson correlations computed among the two meta-indices were statistically significant (both Ps < 0.001). Error bars indicate ± one standard error.

Fig. 2. TMS during the on-judgement phase disrupts metacognition and the response outcome tracking ability of wagered time (WT).

The monkeys demonstrated an impairment in metacognitive efficiency in the TMS-46d condition during the on-judgement phase but not during the on-wagering phase (a). TMS of BA46d does not affect task accuracy (b). Difference in accuracy between unreached trials (low confidence) and reached trials (high confidence) in the on-judgement phase and the on-wagering phase (c, d, respectively). The trendlines are fitted to accuracy by logistic regression with WT as a factor for the TMS-sham and TMS-46d conditions separately. WT reliably tracks response outcomes in the TMS-sham condition but not in the TMS-46d condition during the on-judgement phase. WT tracks response outcomes in both the TMS-sham and TMS-46d conditions during the TMS on-wagering phase (e, f). Distributional differences between correct and incorrect WT. The largest effects were observed in the TMS-sham condition, in which the BA46d was not perturbed (g–j). The WT bin size was set to 1 s; coloured lines indicate kernel density estimation. Error bars indicate ± one standard error; * indicates p < 0.05. ⊗indicates a significant interaction effect (p < 0.05) of WT and TMS (TMS-46d/sham). Shaded areas indicate bootstrap-estimated 95% confidence intervals for the regression estimates.

Fig. 3. On-judgement TMS alters the correlation between reaction time (RT) and wagered time (WT).

No correlation was found between RT and WT in the domain-comparison experiment (a). The Pearson correlation between RT and WT during the on-judgement phase was statistically significant for the TMS-46d condition (p < 0.001) but not significant for the TMS-sham condition (b). The correlations during the on-wagering phase were not significant for either TMS condition (c). RT was significantly negatively correlated with accuracy (correct/incorrect) in the domain-comparison experiment (d) and in both TMS phases in the TMS experiment (e, f). A negative correlation between RT and WT in TMS-46d condition in correct trials (g) but not in incorrect trials (h). Shaded areas indicate bootstrap-estimated 95% confidence intervals for the regression estimates.

Fig. 4. On-judgement TMS distorts the trial-difficulty psychometric curve.

Accuracy decreases with task difficulty (resolution difference; higher values indicate lower task difficulty). The lines are logistic regression fits for accuracy with resolution difference as a factor, calculated separately for the TMS-sham and TMS-46d conditions in the on-judgement phase (a) and on-wagering phase (b). WT decreased with task difficulty in correct trials and increased with task difficulty in incorrect trials in all control conditions (d, f), but this pattern was absent in the on-judgement phase of the TMS-46d condition (c). Distributional differences between TMS-46d and TMS-sham conditions were not significant in either on-judgement phase (g) or on-wagering phase (h), indicating task difficulty were well controlled. Resolution-difference bin size set to 0.02; colored lines indicate kernel density estimation. Shaded areas indicate bootstrap-estimated 95% confidence intervals for the regression estimates.

Fig. 5. Wagered time reflects monkeys’ task performance (correctness) in both memory and perception tasks.

Difference in accuracy between unreached trials and reached trials in the perception (a) and memory tasks (b). Differences between the WTs of correct and incorrect trials for each monkey in the perception (c) and memory tasks (d). WT tracks response outcome (correct/incorrect) in both memory and perception tasks. The lines are logistic regression fits for accuracy with WT as a factor. The WT bin size was set to 1 s; coloured lines indicate kernel density estimation (e). Error bars indicate ± one standard error; * indicates p < 0.05. Shaded areas indicate bootstrap-estimated 95% confidence intervals for the regression estimates.

Fig. 6. Domain-specific metacognition in monkeys.

Task performance in terms of percentage correct was correlated across perceptual and memory domains (a). In contrast, their metacognitive efficiency was not correlated across perceptual and memory domains (b). The DGI quantifies the similarity between their metacognitive efficiency scores in each domain. Greater DGI scores indicate less metacognitive consistency across domains. Darker colours indicate lower metacognitive generality across domains, and the red area indicates the simulated DGI values. The daily domain-generality index (DGI) is shown for each monkey (c) and for all four monkeys (d). The monkeys demonstrate a greater DGI than shuffled data (chance) (e). Two example pairs for pairwise correlation analysis are described (f). The pairwise correlation matrix indicate a pairwise correlation between each monkey and each domain in Hmodel-meta d′/d′ (g) and in accuracy (i). Cluster results from the pairwise correlation matrix in Hmodel-meta d′/d′, revealing two distinct clusters in which data from the same domain grouped together (h), but not in accuracy (j). Error bars indicate ± one standard error; * indicates p < 0.05. Shaded areas indicate bootstrap-estimated 95% confidence intervals for the regression estimates.

Fig. 7. Temporal structure of the TMS experiment.

TMS experiment schedule with TMS-46d/sham conditions counterbalanced between monkeys (Uranus and Neptune) (a). Perceptual judgement task with temporal wagering. Each trial consisted of a starting (blue) cue, a delay lasting 1~6 s, and two simultaneously presented pictures. The monkeys needed to choose the picture with lower resolution (or higher resolution, counterbalanced across monkeys) by holding their hand on the touchscreen. The waiting process was initiated as soon as they laid their hand on the picture. Their confidence in the decision was measured by temporal wagering; that is, they could wait for a reward if they were confident or opt out to abort the current trial. There were two TMS conditions, which differed in the timing of stimulation. In each trial, the monkeys received a single TMS pulse either immediately after the onset of the picture stimulus (on-judgement phase) or 100 ms after they made their perceptual decision (on-wagering phase) (b). The required WT distribution and the actual WT distribution (only catch trials and incorrect trials) with WT bin size set to 1 s. The table depicts the classification of low-confidence trials (unreached trials) and high-confidence trials (reached trials) (c). An illustration of the TMS site, as indicated by the green arrows. Bottom: The green area indicates BA46d on a rendering of a macaque brain; the red disc indicates the target area (d).