Metacognitive ability predicts hippocampal and prefrontal microstructure

The ability to introspectively evaluate our experiences to form accurate metacognitive beliefs, or insight, is an essential component of decision-making. Previous research suggests individuals vary substantially in their level of insight, and that this variation predicts brain volume and function, particularly in the anterior prefrontal cortex (aPFC). However, the neurobiological mechanisms underlying these effects are unclear, as qualitative, macroscopic measures such as brain volume can be related to a variety of microstructural features. Here we used a newly developed, high-resolution (800µm isotropic) multi-parameter mapping technique in 48 healthy individuals to delineate quantitative markers of in vivo histological features underlying metacognitive ability. Specifically, we examined how neuroimaging markers of local grey matter myelination, macromolecular and iron content relate to insight. Extending previous volumetric findings, we found that metacognitive ability, as determined by a signal-detection theoretic model, was positively related to the myelo-architectural integrity of aPFC grey matter. Interestingly, perceptual metacognition predicted decreased macromolecule content coupled with increased iron in the hippocampus and precuneus, areas previously implicated in meta-memory rather than meta-perception. Further, the relationship of hippocampal-precuneus and prefrontal microstructure to an auditory memory measure was respectively mediated or suppressed by metacognitive ability, suggesting a dynamic trade-off between participant’s memory and metacognition. These results point towards a novel understanding of the relationship between memory, brain microstructure, and metacognition. Significance Statement By combining a signal-theoretic model of individual metacognitive ability with state of the art quantitative neuroimaging, our results shed new light on the neurobiological mechanisms underlying introspective insight. Myelination and iron are core determinants of both healthy brain maturation and neurodegeneration; particularly in the hippocampus where iron accumulation is linked to oxidative stress and inflammation. Our results may thus indicate that metacognition depends upon the development and integrity of a memory-related brain network, potentially revealing novel biomarkers of neurodegeneration. These results highlight the power of quantitative mapping to reveal neurobiological correlates of behaviour.


Metacognitive ability predicts hippocampal and prefrontal microstructure 1
Micah Allen1,2, James C. Glen1 Abstract: 10 The ability to introspectively evaluate our experiences to form accurate metacognitive beliefs, 11 or insight, is an essential component of decision-making. Previous research suggests 12 individuals vary substantially in their level of insight, and that this variation predicts brain 13 volume and function, particularly in the anterior prefrontal cortex (aPFC). However, the 14 neurobiological mechanisms underlying these effects are unclear, as qualitative, macroscopic 15 measures such as brain volume can be related to a variety of microstructural features. Here we 16 used a newly developed, high-resolution (800µm isotropic) multi-parameter mapping 17 technique in 48 healthy individuals to delineate quantitative markers of in vivo histological 18 features underlying metacognitive ability. Specifically, we examined how neuroimaging 19 markers of local grey matter myelination, macromolecular and iron content relate to insight. 20 Extending previous volumetric findings, we found that metacognitive ability, as determined by 21 a signal-detection theoretic model, was positively related to the myelo-architectural integrity 22 of aPFC grey matter. Interestingly, perceptual metacognition predicted decreased 23 macromolecule content coupled with increased iron in the hippocampus and precuneus, areas 24 previously implicated in meta-memory rather than meta-perception. Further, the relationship 25 of hippocampal-precuneus and prefrontal microstructure to an auditory memory measure was 26 respectively mediated or suppressed by metacognitive ability, suggesting a dynamic trade-off 27 between participant's memory and metacognition. These results point towards a novel 28 understanding of the relationship between memory, brain microstructure, and metacognition. 29 30 Significance Statement: 31 By combining a signal-theoretic model of individual metacognitive ability with state of the art 32 quantitative neuroimaging, our results shed new light on the neurobiological mechanisms 33 underlying introspective insight. Myelination and iron are core determinants of both healthy 34 brain maturation and neurodegeneration; particularly in the hippocampus where iron 35 accumulation is linked to oxidative stress and inflammation. Our results may thus indicate that 36 metacognition depends upon the development and integrity of a memory-related brain network, 37 potentially revealing novel biomarkers of neurodegeneration. These results highlight the power 38 of quantitative mapping to reveal neurobiological correlates of behaviour. 39 40 SUBMITTED TO JOURNAL OF NEUROSCIENCE 41

Introduction 43
The metacognitive capacity for self-monitoring is at the core of learning and decision-making 44 (Flavell, 1979). As a general capacity, metacognition is thought to enable the flexible 45 monitoring and control of memory, perception and action (Fernandez-Duque et al., 2000). An 46 efficient approach to quantifying this ability lies in the application of signal-detection theory 47 to estimate the sensitivity of self-reported confidence to objective discrimination performance 48 (Fleming and Lau, 2014). Individual differences in metacognitive sensitivity thus quantified 49 are related to the morphological structure, function, and connectivity of the brain (for review, 50 see . Here we expand on these findings using a newly developed 51 multi-parameter mapping (MPM) and voxel-based quantification (VBQ) technique to better 52 elucidate the neurobiological mechanisms underpinning these effects. 53 The anterior prefrontal cortex (aPFC) (Fleming et al., 2010 . While convergent evidence from anatomical, lesion-based, and functional connectivity 58 studies suggest that the right aPFC is specific to perceptual metacognition, metacognition for 59 memory has instead been related to midline cortical (e.g., mPFC and PCC/precuneus) and Although these studies suggest that the ability to introspect on perception and memory depends 62 on the development of a neural mechanism involving both domain-specific and general aspects, 63 the underlying neurobiology driving the relationship between neuroanatomy and 64 metacognition remains unclear. 65 This uncertainty lies partly in the inherent lack of specificity offered by volumetric 66 measures of brain structure, which are fundamentally qualitative in nature. Indeed, voxel-based morphometry (VBM) yields measures in arbitrary units which can be driven by a variety of 68 macroscopic factors such as cortical thickness and variability in cortical folding, owing to a 69 non-specific variety of microstructural features (Ashburner, 2009). It has recently been shown 70 that microstructural properties of brain tissue, such as myelination levels and iron content can 71 lead to the detection of spurious morphological changes (Lorio et al., 2014(Lorio et al., , 2016. The 72 emerging field of in vivo histology aims to combine maps of specific MRI parameters measured 73 via quantitative imaging (qMRI) with biophysical models to provide direct indicators of the 74 microstructural mechanisms driving morphological findings, and ultimately to quantify 75 biologically relevant metrics such as myelination and iron concentrations, oligodendrocyte 76 distributions, and the g-ratio of fibre pathways (Mohammadi et al., 2015;Weiskopf et al., 77 2015). 78 In the present study we used qMRI to map a number of key contrast parameters with 79 differential sensitivity to underlying biological metrics, in order to better understand the 80 microstructural correlates of metacognitive ability. To do so, we acquired high-resolution 81 (800µm isotropic) data using the Multi-Parametric Mapping (MPM) qMRI protocol (Weiskopf 82 et al., 2013). Respecting the quantitative nature of these data, we conducted voxel-based 83 quantification (VBQ) analysis (Draganski et al., 2011) in 48 healthy participants to relate these 84 microstructural markers to individual differences in metacognitive sensitivity during an 85 adaptive visual motion discrimination task. Our results confirmed that right aPFC markers of 86 myelo-architecture positively predict metacognitive ability, whereas left hippocampus and 87 precuneus showed effects consistent with both decreased macromolecule and increased iron 88 content. Further clarifying the domain-general role of memory in metacognition, the 89 relationship of hippocampus and precuneus microstructure with auditory memory was 90 mediated by metacognitive ability. These results extend our understanding of the 91 Task  117   To measure participants' metacognitive ability, we employed a global dot motion  118   discrimination task comprising a forced-choice motion judgement with retrospective  119 confidence ratings on every trial. As part of another investigation, in which we were 120 investigating noise-induced confidence bias (Spence et al., 2015), we used a dual-staircase 121 approach with two conditions in which either mean direction or standard deviation across dot 122 directions was continuously adapted to stabilize discrimination performance. Thus, to control 123 sensory noise independently of task difficulty, in two randomly interleaved conditions we 124 presented either a stimulus with a fixed 15-degree mean angle of motion from vertical and a 125 variable (adaptive) standard deviation (SD), or a variable (adaptive) mean angle from vertical 126 at a fixed 30 degree SD. In either case, the mean (μ-staircase condition, μS) or standard 127 deviation (σ-staircase condition, σS) of motion was continuously adjusted according to a 2-up-128 1 down staircase, which converges on 71% performance. On each trial the motion signal was 129 thus constructed using the formula: Participants were required to judge the global or average motion of a brief dot display, and then 150 rate their confidence in this judgement from 0 (guessing) to 100 (certain). Performance was 151 held constant using an adaptive threshold adjusting either signal mean or variance on each trial 152 (see Methods for more details). Right hand plot demonstrates substantial individual differences 153

Behaviour -Metacognition Global Motion
in metacognitive accuracy, estimated as the type-II area under the curve (AROC), 154 independently of motion discrimination performance. Inter-individual differences in AROC 155 were then used in a multiple regression analysis to explain variation in microstructural brain 156 features (see VBQ Analysis). 157 158 Participants were instructed that the goal of the task was to measure their perceptual and 159 metacognitive ability. Metacognitive ability was defined as a participant's insight into the 160 correctness of their motion judgements, i.e. how well their confidence reports reflected their 161 discrimination accuracy. Participants completed a short practice block of 56 trials, in which 162 they performed the motion discrimination without confidence ratings, with choice accuracy 163 feedback provided by changing the colour of the fixation to green or red. All participants 164 achieved better than 70% accuracy and indicated full understanding of the task before 165 continuing. Participants completed 320 trials divided evenly between the two staircase 166 conditions. Trials were divided into 10 blocks each with 40 trials, randomly interleaved across 167 conditions within each block. 14 participants did not complete the last two blocks of the task 168 due to a technical error, however all participants had at least 100 trials per condition (Fleming 169 and Lau, 2014).

Data Acquisition 192
All imaging data were collected on a 3T whole body MR system (Magnetom TIM Trio,  193 Siemens Healthcare, Erlangen, Germany) using the body coil for radio-frequency (RF) 194 transmission and a standard 32-channel RF head coil for reception. A whole-brain quantitative variable. Importantly, we followed previous investigations and controlled all analyses for 278 average discrimination sensitivity (d-prime), confidence, response bias, the variance-induced 279 confidence bias, and the difference in mean signal between the two staircases. To estimate 280 variance-induced confidence bias, we fit multiple regression models within each subject, 281 modelling trial-wise mean, variance, accuracy, and RT as predictors of confidence. This 282 provided beta-weights for each participant indicating the degree to which their confidence 283 report reflected variance-specific bias independently of the other modelled factors, which were 284 then included in our VBQ multiple regression. 285 286 Following standard VBM procedure, we also included age, gender, and total intracranial 287 volume as nuisance covariates. We then conducted small-volume corrected analyses of the 288 positive and negative main effect of metacognitive ability (AROC) within our a priori mask, 289 correcting for multiple comparisons using Gaussian Random Field Theory, FWE-peak 290 corrected alpha = 0.05. Further, we analysed the whole-brain maps of the same contrasts, using 291 a non-stationarity corrected FWE-cluster p-value with a p < 0.001 inclusion threshold 292 (Ridgway et al., 2008;Hupé, 2015). All anatomical labels and % activations were determined 293 using the SPM Anatomy Toolbox (Eickhoff et al., 2005). 294 295

Mediation Analysis of Auditory Memory and Metacognition 296
To explore the relationship of metacognition, memory, and brain microstructure we conducted 297 a single-level mediation analysis with the auditory memory score (X) predicting brain

Behavioural Results 317
To check staircase stability, we first performed two-way repeated measures ANOVA (factor 318 A: block, levels 1-7; Factor B: staircase condition, μS vs σS) on accuracy scores after removing 319 the first block. As several subjects did not complete the last two blocks, we first re-binned trials 320 into 8 equal size bins of 20% total trial length, before analysing block stability. This analysis 321 revealed a significant main effect of variance on accuracy (F(1, 47) = 15.15, p < 0.001), but no 322 main effect of block (ps > .33) or block by condition interaction (p > 0.11), indicating that 323 although average performance was slightly higher in the σS condition (Mean Accuracy μS = 324 73.4%, Mean Accuracy σS = 76.6%), this difference did not change over time, indicating stable 325 performance within each staircase. As a further check we repeated this analysis separately 326 within each condition; in both cases the block main effect was not significant (all ps > 0.13). 327 All participants thus achieved stable performance, with an average accuracy of 75.3% (SD = 328 3%) across two conditions. Metacognitive ability was comparable to previous studies using the 329 AROC (mean AROC = 0.68, SD = 0.06) and did not differ between conditions, t (47) = -0.67, 330 p = 0.51. Table 1

Whole-brain AROC Analysis
Our whole brain analysis revealed a striking relationship between AROC and left hippocampal 362 MT. Here, higher AROC related to reduced MT in the left posterior-hippocampus (peak voxel 363 MNIxyz = [-31 -25 -14]). Inspection of this result in the SPM anatomy toolbox revealed that the 364 majority (51.6%) was in the dentate gyrus (33.4% 'activated'), with another 29.3% in CA1 365 (14.2% activated), and to a lesser extent in the subiculum (6.6%, 1.8% activated) and CA3 366 (5.6%, 14.9% activated). Additionally, we found that iron levels as indexed by R2* negatively 367  In contrast, B) the relationship of memory and hippocampal-precuneus (HP) myelo-415 architecture is supressed by metacognitive ability (blue arrow), but enhanced for HP iron (as 416 measured by R2*). These results suggest individual differences in memory and metacognition 417 are related by a dynamic interaction between the two networks, with the memory-related 418 network increasingly related to metacognitive circuits and vice versa for the prefrontal cortex. 419 Statistical significance for each parameter of the path model (a, b, ab, and c'; * < 0.05, *** < 420 0.001) determined via bootstrapping procedure, see Methods for more details. 421 422 Discussion 423 Our findings demonstrate that individual differences in metacognitive ability are related to 424 underlying microstructural features of prefrontal and hippocampal neuroanatomy. Previous 425 studies investigating individual metacognitive ability indicated that the function, connectivity, 426 and volume of anterior prefrontal cortex (aPFC) underlie introspective insight. Here we build on these findings using a novel quantitative magnetic resonance technique to show that 428 prefrontal correlates of metacognition are related to markers of grey-matter myelo-architecture, 429 as indexed by the overlapping effect in both MT and R1 maps. In contrast, we found that 430 differences consistent with decreased hippocampus and precuneus macromolecular content, 431 coupled with increased iron content predicted metacognitive ability. Interestingly, we also 432 found that metacognition mediated the relationship of memory ability and microstructure in 433 the hippocampus and prefrontal cortex, suggesting a domain-general mechanism linking 434 memory and metacognition. These results suggest that unique neurobiological mechanisms 435 underlie the development and aetiology of metacognition for memory (i.e., meta-memory) and 436 perception (meta-perception). 437 Previous investigations of metacognitive ability and individual differences in brain 438 anatomy suggest that introspection for memory and perception depend upon both common and 439 unique neural substrates. As strong evidence for their dissociation, Fleming et al (2014)  440 recently demonstrated that medial-prefrontal (MPFC) or aPFC lesions selectively disrupt meta-441 memory or perception, respectively. In another study by McCurdy et al (2013), although 442 metacognition for perception and memory were found to correlate, the two processes 443 independently related to the volume of aPFC and precuneus. Interestingly, in an interaction 444 analysis these authors found that while precuneus volume also predicted meta-perceptual 445 ability, aPFC did not predict meta-memory. In contrast, Baird and colleagues (2013) found no 446 behavioural correlation of meta-memory and perception, but instead found that the former 447 predicted functional connectivity from the medial prefrontal cortex (mPFC) to the precuneus 448 and inferior parietal cortex, while the latter was predicted by functional connectivity seeded 449 from the aPFC to the cingulate, putamen, and thalamus. However, Baird et al. also examined 450 the differential connectivity strength of aPFC vs mPFC and found that better meta-memory 451 was associated with stronger connectivity between the aPFC and the hippocampus, precuneus, and other memory-related areas, a finding which may possibly explain our mediation results. 453 Thus while these studies suggest a degree of independence between meta-memory and meta-454 perceptual systems, they also suggest that interactions between memory (precuneus, 455 hippocampus) and metacognition-related (aPFC) brain areas underlie general metacognitive 456 ability. 457 Complementing these results, we found that perceptual metacognitive ability was Of course, the causal direction and precise mechanisms underlying these effects remain 508 unclear; it could be for example that individuals with a less stressful environment have a better 509 overall memory (and therefore more evidence of neuroprotective element) (Rodrigue et al.,510 2013), or the iron profiles indicated here may represent a direct response to stress and/or reflect 511 a neuro-genetic ability to adaptively respond to stress or nutritional challenges. We might 512 speculate that metacognition-related hippocampal histology could potentially be an early 513 biomarker for Alzheimer's and other neurodegenerative diseases; if such a pattern does reflect 514 an early-life response to stress, it may come at the cost of an increased risk to developing these 515 diseases later in life. Regardless, our study provides an interesting starting point for future 516 work, so that we can better understand how the histological factors discovered here relate to 517 metacognitive ability. 518

Limitations and Future Directions 519
Methodological limitations common to all studies requiring spatial normalisation are the 520 potential for residual registration errors, as well as partial volume effects. To minimise these 521 sources of bias we used the DARTEL algorithm for inter-subject registration, which results in 522 maximally accurate registration (Klein et al., 2009), and used the voxel based quantification 523 normalisation procedure to minimise partial volume effects introduced by smoothing 524 (Draganski et al., 2011). 525 Here we speculatively interpret our results as suggesting a link between the histology 526 of the hippocampus and APFC, and the relationship of memory and metacognition. Indeed, the hippocampus and precuneus are a core part of a memory-related network (Squire, 1992;Brown 528 and Aggleton, 2001). We also found that metacognitive ability both negatively mediated 529 (supressed) the relationship of auditory memory ability and hippocampal/precuneual 530 microstructure, and positively mediated the link between memory and APFC myelination. This 531 suggests that a dynamic interaction between memory and meta-cognition related brain areas 532 mediates the influence of metacognition on memory (or vice versa); however future work with 533 a more general battery of memory and meta-memory tasks is needed to better understand this 534 relationship. Additionally, without stress or nutrition-related measures we can only speculate 535 as to their link with the effects observed here. Indeed, individual differences in brain structure 536 and function are influenced by a variety of developmental, neurogenetic, and environmental 537 factors (Kanai and Rees, 2011). 538 Thus, an important step for future studies will be to study the microstructural correlates 539 of metacognitive ability with a more comprehensive battery of perceptual, memory, and stress-540 related measures. Nevertheless, our study is the first to establish that quantitative neuroimaging 541 reveals potential biomarkers sensitive to metacognitive ability, which has allowed us to greatly 542 extend previous functional and volumetric studies. We therefore anticipate that future research 543 will expand upon these findings to better understand their developmental, genetic, and 544 environmental causes. 545 546

Conclusion 547
By using a quantitative multi-parameter mapping approach, we were able to reveal the 548 contribution of the hippocampus and precuneus to perceptual metacognition. Furthermore, as 549 our method yields standardised quantitative metrics sensitive to the underlying tissue 550 microstructure, these results can be used to inform future clinical research as they can be 551 compared directly across research sites. Our results suggest that the concentration of iron in the hippocampus is an important predictor for metacognitive ability, pointing towards a 553 potential neuroprotective mechanism defending introspection from stress. More generally, 554 these results suggest that memory-related systems may be more important than previously 555 realized for the computation of perceptual confidence. Future research into the genetic, 556 environmental, and other dynamic factors underlying these findings are likely to yield strong 557 dividends in the study of metacognitive ability.  peak-corrected result and pvalues. ‡ indicates result from exploratory p < 0.01 inclusion threshold. k = cluster 572 extent in voxels, T = t-value, Z = z-value, x,y,z = MNI peak coordinates. pFWE cluster = family-wise corrected 573 cluster p-value, pU cluster = uncorrected cluster p-value, pFWE peak = family-wise corrected peak p-value, pU 574 cluster = uncorrected peak p-value. + -indicates positive or negative t-constrast. See VBQ analysis and VBQ

575
Results for more details.