Additionally, they had to describe IV the content of their thoughts between trials, and V if they noticed any changes in their performance during the experiment. In order to minimise signal dropout due to the frontal sinuses, the slices were tilted away from the coronal orientation by an angle of Particularly for ultra-high field strength, signal dropout and distortions in frontopolar cortex can be substantial.
This setup was found to maximally reduce signal distortions and dropouts for the present study because it allowed us to use a small field of view FOV , and thus a short echo train length, in order to cover most of anterior prefrontal cortex with maximal exclusion of the air-filled cavities compared to axial slices Fig. However, using this setup no region beyond frontopolar cortex could be covered. Data was acquired for 10 functional runs, each lasting 5 minutes volumes per run.
The first two volumes of each run were discarded by default to allow for magnetic saturation effects. Subjects lay supine in the scanner and viewed the projection via a mirror. Responses were recorded using a set of two custom-engineered, deconstructed Nintendo Wii joysticks, each with two buttons operated with the index- and middle fingers of either hand. A Example of one slice of one participant's EPI image. B Structural T1 image from the same subject displaying the positioning of the example slice dotted line and slice coverage blue box.
Due to the optimized slice positioning, which allowed the use of a small field of view FOV and a short echo train length, a relative small part of the air-filled cavities was included. This improved the quality of the EPIs and reduced signal dropouts and distortions. C The parameter estimates from the FIR model were used for multivariate pattern classification. Multivariate pattern classification was used to analyse the data. The analysis sought to identify regions within FPC that allowed subjects' decisions for left and right to be decoded from fine-grained patterns of activity as measured by the BOLD signal preceding the subjects' conscious awareness of their decisions.
No additional normalization or smoothing was performed at that stage in order to maximize the sensitivity for information encoded in the fine-grained voxel patterns  ,  , . A finite impulse response FIR predictor was used to model fMRI responses, as it was not known whether the profile of the fMRI time course adheres to the generic haemodynamic response function in this situation. This procedure also allowed time-resolved decoding to be implemented  , . Left-button trials and right-button trials were modelled as two separate conditions, each with 20 FIR regressors.
Each regressor modelled a time-bin of 1. The 10 th time-bin was defined as that in which a decision was made. The first 10 regressors therefore modelled the 15 seconds preceding and including each decision, the last 10 covered the 15 seconds following the decision. Invalid trials in which subjects were unable to recall a letter were modelled separately, again using 20 FIR predictors and assigning the 10 th predictor as that in which the button was pressed.
These trials were excluded from the pattern classification analyses. To minimise unaccounted-for variance in the fMRI data, the second button-presses with which subjects indicated the letter present at the decision time were modelled as covariates. Left-handed and right-handed button-presses were modelled separately, and convolved with a standard haemodynamic response function HRF.
This way the choice-predictive information encoded in a spherical cluster of voxels at each position in the brain can be estimated without making any a-priori assumptions as to the location of the information Fig. As a direct replication of Soon et al. Nevertheless, information decoding was based on the same number of data points voxels in both cases. For every voxel in the volume denoted by v i , a spherical searchlight cluster of N voxels was defined around it, such that all voxels in the cluster denoted by c 1…N lay within a radius of 3 voxels of v i. For every voxel in c 1…N , the parameter estimates for all 20 time-bins were extracted from each run, separately from left-decision and right-decision trials.
These were transformed into two N-dimensional pattern vectors one corresponding to left-decision trials, the other to right-decision trials for each of the 20 time-bins, representing the spatial activation patterns for both decisions at all 20 points in time. The classifier estimated a decision boundary separating the two classes of patterns in N-dimensional space where N is the number of voxels in the local spherical cluster also see .
This procedure was repeated 10 times, each time using a different run as the independent test data set, resulting in a fold cross-validation. The pattern classification results from each repetition were averaged and assigned to the central voxel of the searchlight cluster as its decoding accuracy. The entire procedure was repeated by assigning in turn every voxel in the brain volume as the central voxel of the searchlight cluster, yielding a 3D map of decoding accuracies throughout the imaged volume. Furthermore, such a 3D decoding accuracy map was obtained for each of the 20 time-points.
These maps represent the amount of intention-related information encoded in the local neural networks at each location in the brain of each individual subject , at the time-point from which the parameter estimates were taken. In the next step, the subjects' individual decoding accuracy maps were normalized to MNI-space. For this, the functional images were first co-registered to the individual high-resolution T 1 -weighted structural whole-brain image, which was acquired during the same scanning session.
The normalization parameters were then applied to the decoding accuracy maps. Voxels that were not shared by all subjects were masked out. For each time-point, group level analyses were performed across subjects. It was therefore possible to track changes in the amount of information encoded in different regions over time, and in particular, to search for a build-up of intention-related information prior to subjects' conscious awareness of their own intentions, as observed by Soon et al.
The goal of this analysis was to investigate the spatial-temporal profile  of the time-bins that allowed the prediction of free decisions before they reach conscious awareness. Individual data from the searchlight yielding the highest decoding accuracy across subjects see Results and Fig. For this, this coordinate, which was established after normalizing the individual 3-dimensional decoding accuracy maps, was transformed back into individual space. FPC only showed significant decoding accuracies in the time-bins preceding the decision.
B The graph displays the average time-course of decoding accuracies, taken from the central voxel of the searchlight cluster that showed the highest decoding accuracy. Error bars represent standard errors. The time-bin of the conscious intention is indicated by the red bar and is labelled as time 0. One time-bin corresponds to 1. Coordinates displayed are MNI coordinates. The pattern vectors from single time-bins were then combined for each decision by I simply averaging vectors in steps of i two, ii three or iii four time-bins or by II concatenating vectors in the same steps.
The multivariate pattern classification analysis was run again on these new vectors, exactly as described before; the difference was that there were a smaller number of time steps per analysis. If averaging across earlier time-bins does not reduce the decoding accuracy, this would mean the spatial activation patterns display a consistently high temporal stability in those time-bins.
Finally, correlation analyses were also conducted between the pattern vectors of adjacent time-bins separately for each subject and each condition , in order to assess the temporal stability of these patterns in more detail. Since the results did not differ for left and right decisions, the results are reported for each time-bin for all the analyses.
Please note that all these subsequent analyses only aimed to specify the role of the best searchlight cluster and not to select voxels for further dependent statistical analyses, which would have been circular . Also note that the chosen cluster was the best decoding cluster averaged across subjects. This cluster therefore did not represent the optimal decoding cluster in individual subjects. Analysing the optimal clusters in individual subjects, however, would have carried the risk of arbitrariness and was therefore strictly avoided.
The parameter estimates obtained from a GLM, based on normalised and smoothed 3 mm FWHM data, were used in a conventional mass-univariate analysis. Again, a finite impulse response FIR predictor was used to model fMRI responses identical to the analysis described above and group level analyses were performed across subjects for each time-point separately. The purpose of this analysis was to investigate whether any voxels at any time-point showed significant differences in activation between the left-decision and right-decision trials. In nearly all trials, subjects indicated that the decision reached conscious awareness during the presentation of the same letter or one letter before they pressed the button see Table 1 for details.
One subject S4 was thus excluded.
For nine of the remaining eleven subjects, the individual probability of left and right decisions did not significantly differ. This was even longer compared to the original study  ; spill-over effects from the previous trial therefore can not explain the results. In the post-experimental interviews subjects indicated that they were able to relax and make spontaneous decisions.
Most subjects reported that they did not have specific thoughts they could remember. Some reported having thought about or mentally read the letters, some reported having occasionally thought about daily activities but none reported having thought about the decisions. Most subjects reported that they became more relaxed through the experiment and that they either became more spontaneous or that there was no change in spontaneity. This is not surprising given that subjects were highly familiar with the task, having completed 10 runs of prior training, and were able to perform the task effortlessly.
Comparable to the original study  , the distribution of sequence lengths periods of the same decision before a switch occurred resembled an exponential distribution, as expected for random behaviour. Furthermore, we correlated the sequences of decisions from each run of each subject with the sequence of decisions in the following run, in order to control for the possibility that subjects might have simply repeated fixed sequences of decisions over the experiment.
For each subject and within each functional run, we further analysed whether the sequences of left and right decisions violated the assumption of a random order runs test as implemented in MATLAB, MathWorks Inc. We note, however, that due to the nature of our task, the number of successive trials per functional run were very short, thus, this test has limited informative value. These results indicate that participants performed correctly and that preplanning or other unaccounted-for activity cannot explain the decoding results see Table 2 for individual results.
Multivariate pattern classification analysis was used to search for brain regions encoding subjects' decision outcomes. We identified a cluster in FPC from which subjects' decisions could be decoded before their intentions became conscious i. Taking into account the temporal delay of the BOLD signal which is in the order of a few seconds , it is possible that these signals reflect processes up to 10 seconds before the actual decision. Grey transparent voxels did not show decision preference or were not located in grey matter.
Colours are scaled for better visualization. Informative patterns were different for each participant. When the radius was further increased, no significant results could be achieved, possibly due to the increased dimensionality. In a control analysis the accuracy maps from the time-bins after the decision was made time-bins 11—20 were contrasted against chance level. No clusters could be found in FPC encoding any information above chance level during this period. The same held true if separate time-bins around the time of the motor response were considered, matching the findings from the original study that FPC only encoded the intentions before subjects were aware of making a decision.
The information was, as in the original study, only encoded in fine-grained activation patterns rather than in the average signal. This held true for the whole frontopolar region as well as for the region from which decoding was possible. A Histogram of sequence length. Displayed is the average percentage of sequences of N trials of the same decision left or right before switching to the other decision. This suggests that subjects made random decisions.
Error bars are standard errors. For both conditions, the signal increased only after the decision red bar and came back down to baseline in the next ten seconds. Significant differences between left and right decisions were not found for this cluster. Coordinates are given as MNI coordinates. The subsequent decoding analysis using the best searchlight cluster across all subjects showed that averaging across adjacent time-bins led to lower decoding accuracies, the more time-bins that were combined Fig. Concatenating time-bins, which is combining spatial and temporal information before the decision, still predicted the decision outcome with high accuracy Fig.
Concatenating was superior to averaging by trend, which suggests that, in the time leading up to a decision, spatial patterns were not uniform throughout, but carried more decision-related information with increasing temporal proximity to the decision. In subsequent correlation analyses we found that patterns from consecutive time-bins nevertheless showed significant correlation. After the time-point of the decision, the correlations dropped again to a stable level.
This auto-correlation curve closely mimicked the time-course of decoding accuracies Fig. Thus, activity patterns became more similar and more informative the closer the decision-maker was to becoming aware of the decision. After the decision was made, some pattern stability was sustained but the patterns no longer carried information about the decision. Temporal-spatial decoding analysis. The reference time-bin for vector concatenation was the time point of the decision time 0 s. The resulting pattern vectors additionally represented temporal information for the best searchlight cluster and were used for multivariate decoding.
Temporal-spatial information was found to be highest directly preceding the decision and was still present when four time bins were concatenated. Concatenating was superior to averaging by trend. Displayed is the decoding accuracy across time from the best cluster empty gray triangles as well as the correlation of each time-bin with its preceding time-bin filled yellow triangles as a measure of pattern similarity averaged across patterns for left and right decisions. Up to the time of the decision time 0 s the decoding accuracy and pattern similarity increased in a similar fashion.
After the decision, the pattern similarity dropped slightly and patterns did not predict the decision outcome anymore. This study aimed to assess whether local spatial activity patterns in FPC, which were previously found to encode unconscious intentions  , display temporal stability over time. We could also demonstrate that activity patterns preceding the time-point of the conscious decision became increasingly similar with increasing temporal proximity to the decision.
Our behavioural data and questionnaire results further supported that no conscious processes biased the decisions. Thus, early predictive activity patterns are attributable to unconscious components of evolving intentions. Comparable to the original study  subjects' intentions could be read out approximately seven seconds before they became conscious. Given the haemodynamic delay, it is likely that this reflects neural processes that occurred even earlier by a few seconds. The site of information encoding was found to be left frontopolar cortex, also referred to as the rostral lateral prefrontal cortex or the anterior prefrontal cortex, and approximating to the most anterior part of Brodmann area 10  — .
The same region was identified in the original study but in the opposite hemisphere. In the present study, we optimized the slice positioning to minimize distortion effects and signal dropouts, which are a common problem due to the proximity to frontal sinuses, especially at higher field strength. Since the analysis only included voxels that were present in all subjects, residual dropout in individual subjects could have led the exclusion of more informative voxels.
Hence, our results might underestimate the extent of the decision-related region. No information about the subjects' intentions was found after the decision was made, which is also in line with the original findings that after the time of the decision, information was only encoded in primary motor cortex and pre-motor cortex .
These areas were not covered in the present study. As demonstrated before, the procedure used in both studies ensured that decoding could not be explained by activity related to the previous trial. As in our previous work  ,  we used a Finite Impulse Response FIR model, which is designed to separate effects of the current trial from the previous and the following trial. This method is highly efficient as long as both types of responses are roughly equally frequent, as here. Importantly, subjects self-paced their decisions, ensuring that the intervals between trails were variable, which makes the estimation of the FIR model even more robust to carry-over effects.
The average trial duration in the present study was even longer than in the original experiment, thus making it less likely that spill-over effects from the previous trials might have occurred. For the earliest time points in a trial we find no predictive information as would have been expected if carry-over effects occurred. However, as time continues, we begin to see information. Third, the temporal resolution was also improved 1.
Fourth, data from one trial cannot be used to predict the trial preceding or following it. Fifth, carry-over effects are further unlikely because the distribution of response sequences resembled an exponential distribution, as expected for random behaviour Fig.
Although we do not believe and do not claim that our subjects produced perfectly random sequences, our behavioral results suggest that subjects made spontaneous decisions. This is probably because we did not ask subjects to balance their decisions. Interestingly, we observed an increase in similarity between patterns with increasing temporal proximity to the conscious decision. This increase in correlation was mirrored by the increase in information content about the decision outcome.
This hypothesis states that once a threshold is crossed a certain pattern is stable enough , a conscious decision is made and activation patterns lose their predictive power afterwards. The remaining but reduced pattern stability might be explained by the dependence of sequentially acquired brain scans. Although there was some tendency for patterns to remain stable for a few seconds after the decision, there was no decodable information at these post-decision time periods.
Similarly, patterns during the initial phase of the following trial were not informative. It was only later, closer to the next decision in the next trial, that we again observed a slow increase of pattern similarity and information encoding. This again speaks against carry-over effects from the previous trial. Our detailed behavioural analysis confirmed that subjects did not use any systematic thoughts to consciously prepare their decision ahead of time.
They acted as instructed and were spontaneous. In support of these findings, Soon et al. Here this analysis was not possible due to the restriction of coverage to PFC which was necessary to achieve a higher spatial resolution. We thus conclude that the early informative spatial activation patterns in frontopolar cortex were related to unconscious components of the intention.
It might be surprising that decision-related information is encoded in the brain several seconds before the decision becomes conscious, given that the task was rather simple. One possibility is that random activity directly preceding the decision might bias the decision outcome, as suggested for short time periods . This, however, is less likely for such long periods as observed here. Our study might have facilitated the detection of very early information by encouraging subjects to relax and refrain from decision-related thoughts as well as by instructing subjects to self-pace their decisions.
By doing so, unlike most other studies, our experiment was uniquely suited to investigate the early evolution of intentions. Another possibility is that, even though we have good evidence that our subjects' behaviour was spontaneous, there might still have been some urge to respond regularly to a certain degree, which might only become detectable in longer behavioural sequences than produced here. Such a bias, even though outside subjects' awareness, could potentially contribute to the build-up of early brain activation patterns.
It is important to note that any temporal autocorrelation in the signals could cause a correlation between choices in successive trials, even without a conscious, deliberative link. Such autocorrelation might be considered a very basic form of memory, but our conclusion that choices can be predicted before awareness would remain unchanged.
Please also note that our study cannot provide evidence for a causal relationship between the activation in frontopolar cortex and the decision, e. Our study also did not address the question of whether inter-individual differences, e. Thus, one intriguing question for future research will be whether the onset of unconscious decision formation that can be decoded from brain activity might in turn be predictable by some core cognitive ability.
The present study supports the hypothesis that prefrontal cortex is a core region for free decisions. Presently, it is believed that the anterior prefrontal cortex lies at the top of a hierarchically organized prefrontal functional architecture. Prefrontal cortex represents sensory input information in its most abstract form and guides cognitive control . It maintains the abstract representation of a desired act, together with context-relevant information such as environmental context, task-rules, motivation and potential outcomes  ,  —  ; the motor plan for the execution of this act is prepared in premotor areas; this is broken down into co-ordinated recruitment of single motor units in primary motor cortex .
Medial prefrontal cortex might additionally contribute to action planning by processing self-related information  , in this case, one's intentions; it was also found to encode freely chosen decisions during a delay . Of the different regions in prefrontal cortex, however, evidence from cytoarchitectural studies suggests that frontopolar cortex has the necessary architecture to support the highest level of processing within prefrontal cortex.
First, it has the greatest number of dentritic spines per cell, and overall spine density is higher than for all other areas of prefrontal cortex. Furthermore, it is the only supramodal area that is connected solely with other supramodal areas, has less laminar differentiation compared to other prefrontal areas, and its connections within PFC point towards a hierarchically high level of processing  ,  , . Given these properties, frontopolar cortex is an optimal candidate for the representation of the most abstract contents . Current hypotheses about the function of this region are based mainly on functional imaging studies as this region is markedly smaller and difficult to access in primate electrophysiology.
Presently, the cognitive processes in which frontopolar cortex has been implicated include: processing of internal states  , modulation of episodic memory retrieval  ,  , prospective memory  , relational reasoning  ,  , the integration of cognitive processes  and cognitive branching . Intuitively, we would therefore expect to see phenotypic differences across all ranges of the polygenic scores and more acutely with the extremes of the distribution. The most comprehensive study to have examined the association between polygenic scores for AD with cognitive function [ 28 ] used a predictor based on the Lambert et al.
The independent target dataset in that study was the UK Biobank study. Small but significant associations, not explaining more than 0. In conclusion, there is potential clinical utility for the stratification of mid-to-late-life population-based cohorts into high and low risk groups based on APOE status and global polygenic risk to better understand the pathophysiology of AD.
However, large sample sizes for both the GWASs used to build the polygenic scores and to select at risk sub-groups of the population are likely to be necessary. However, the extremes of the polygenic score distribution will be of additional value as, by definition of their construction, they will tap into genome wide risk and multiple pathways that lead to AD.
This work was supported by multiple sources. We are grateful to all the families who took part, the general practitioners and the Scottish School of Primary Care for their help in recruiting them, and the whole Generation Scotland team, which includes interviewers, computer and laboratory technicians, clerical workers, research scientists, volunteers, managers, receptionists, healthcare assistants and nurses. We thank Professor Julie Williams for providing comments on an early draft of the manuscript. Alzheimers Dement 7, — Lancet Psychiatry 3, — BMJ Open 2, pii: e Lancet Neurol 15, — Nat Rev Neurol 9, — Annu Rev Neurosci 37, 79— Nat Genet 7, — Am J Phys Anthropol , — PLoS One 9, e Nature , — Brain Pt12 , — Eur J Hum Genet 24, — The study, its participants and their potential for genetic research on health and illness.
Int J Epidemiol 42, — BMC Med Genet 7, Lezak MD Neuropsychological Assessment, 3rd edition. Gigascience 4, 7. The Genomes Project, Consortium A map of human genome variation from population-scale sequencing. Am J Hum Genet 88, 76— BMC Genomics 16, Lancet , — BMC Med Genet 14, Bioinformatics 31, — Nat Genet 45, — Mol Psychiatry 21, — R package version 1. R package version 2. Alzheimer Dis Assoc Disord 30, — Nat Genet 47, — Curr Biol 26, — Nat Neurosci 17, — Flowchart documenting the selection process of the Generation Scotland analysis cohort.
All models adjust for age, sex, and pedigree-based relatedness.
Statistical analyses Linear mixed modelling was used to test for differences in cognitive ability by AD polygenic risk scores and APOE status. Sensitivity and secondary analyses While a kinship matrix was included to model relatedness between participants, a sensitivity analysis on only unrelated individuals was performed.
Figures and Tables Fig. Show more.