Abstract

Results

We investigated the neural precursors of abstract intentions by asking participants to perform an abstract decision task in which they spontaneously and freely chose to perform either of two mental arithmetic tasks: adding or subtracting (Fig. 1 and Experimental Procedures). Participants were presented with a continuous stream of stimulus frames, each consisting of a letter and five single-digit numbers. They were asked to stay relaxed while monitoring the letter stream, which served essentially as a “clock” for reporting the timing of their decision. Immediately when they felt the spontaneous urge to perform either addition or subtraction, they first had to memorize the letter visible on the screen, which allowed us to reconstruct the timing of their free choice. Then they had to immediately perform the freely chosen task on the two numbers that were subsequently presented in the center of the two following frames. The third frame from the time of decision served as a response mapping screen for the calculation task, with four possible response options in the four corners: the correct addition answer, the correct subtraction answer, and two incorrect response options. The positions of these answers were randomized from trial to trial, and participants had to press a button to indicate their choice, thereby also revealing which calculation task they had chosen. After this first response, a second response-mapping screen consisting of four letters was displayed and participants indicated which of the letters was visible on the screen when they consciously made their decision. Seventeen participants performed the task while undergoing functional magnetic resonance imaging (fMRI) at 3 T, with one functional scan acquired every 2,000 ms. To identify any neural precursors that may be shaping the upcoming abstract intentions, multivariate pattern classification (18–22) was used in separate analyses to look for brain activity that could predict the content and timing of the abstract decision.

Open in a separate window

Fig. 1.

Measuring the onset and content of spontaneous abstract intentions. A trial began with a continuous series of stimulus frames refreshed every second, each consisting of a central fixation point, a letter below it, a single-digit number above it, and four single-digit response options, one in each corner. Immediately when participants felt the spontaneous urge to perform either adding or subtracting, they first noted the letter on the screen (frame 0 relative to time of decision). The chosen arithmetic task was then performed on the numbers presented above the central fixation in the next two stimulus frames (frames 1 and 2). The response options for the numbers in frames 1 and 2 were randomly presented in the four corners of the subsequent stimulus frame (frame 3): the correct addition answer, the correct subtraction answer, and two incorrect response options. Participants selected the correct answer for the chosen task by pressing one of four corresponding buttons, thereby revealing the content of their abstract decision. After the response was given, four letter options were presented from which participants selected the letter presented at frame 0, thereby revealing the time of conscious decision.

It was stressed to participants that the decision time and choice of task were completely up to them but that they should be as spontaneous as possible and execute their decision without hesitation once it was made. We ensured that all mathematical problems and their correct answers consisted of only single-digit numbers, so either task could be easily performed within the short duration given without needing much effort or preparation. Participants were familiarized with the two tasks, having gone through a prior behavioral training session. Only participants who were highly proficient in both tasks were selected for scanning. In the fMRI experiment, the inappropriate numbers were only selected on 1.2% (SE = 0.3%) of the trials, indicating that participants performed the chosen task correctly instead of making random responses. The participants’ decisions could thus be reliably inferred, based on the choice of one of the two correct answers. For the appropriate responses, participants randomly (Fig. S1) selected between adding (51.9% of trials; SE = 2.0%) and subtracting (48.1%; SE = 2.0%) with equal probability (t₁₆ < 1). Given that the four response options were randomly arranged, the specific motor response was uncorrelated to the choice of task. Therefore, any information regarding the content of the decision that was decodable from brain activity could not possibly be a result of covert motor preparation. In 97.4% (SE = 0.5%) of the trials, participants indicated that the decision was made three frames before the response; that is, in frame 0, showing that they followed the task execution sequence exactly as instructed. On average, participants made a spontaneous decision 17.8 s (SE = 1.8 s) after trial onset, resulting in 12.2 (SE = 0.8) decisions per experimental run. This long delay facilitated the search for unconscious neural precursors of the decision while avoiding contamination from the preceding trial.

We first investigated which cortical regions contained predictive information about the outcome of the abstract decision to perform addition or subtraction, and whether such information was available before or after the decision reached conscious awareness. For each discrete point from 8 s before to 18 s after the decision (14 points, each 2 s), and in different regions across the whole brain, independent classifiers were trained to distinguish between the spatial patterns of brain activity related to the two choices (adding and subtracting). The accuracy with which each independent classifier could predict the specific choice revealed whether a particular cortical region contained information related to the content of the intention at a specific point in time. This method has previously been shown to be highly sensitive in decoding simple motor intentions and in uncovering the flow of information in task preparation (2, 23).

We found that up to 4 s before the conscious decision, a medial frontopolar region (P < 0.00005 uncorrected, 5-voxel cluster threshold, 59.5% accuracy) and a region straddling the precuneus and posterior cingulate (P < 0.00005 uncorrected, 5-voxel cluster threshold, 59.0% accuracy) began to encode the outcome of the upcoming decision (Fig. 2). During this early phase, the overall signal in both regions did not show any significant change from baseline (t₁₆ < 1), nor was there any significant difference between addition trials and subtraction trials (t₁₆ < 1), suggesting that the information was encoded in the fine-grained spatial pattern of activation, rather than any global increase or decrease in neural activity (Fig. S2). We ensured that this early information was not a result of carry-over of information from the previous trial (SI Text S1).

Open in a separate window

Fig. 2.

Decoding the outcome of abstract decisions before and after they reach conscious awareness. Projected onto the medial cortical surface are brain regions that predicted the outcome (red) of the abstract decision before it was consciously made (MNI coordinates). Inset shows similar results for the decoding of free motor decisions before conscious awareness in our previous study (2). The lateral surface shows the region that encoded the outcome of the decision after it became conscious. Line graphs depict for each cortical region the accuracy with which the abstract decision to perform addition or subtraction could be decoded at each time (error bars, SE; chance level, 50%). The vertical red line indicates the point of conscious decision, and the vertical gray dashed line indicates the onset of the next trial. Given the hemodynamic delay, information available at 0 s would have been a result of neural activity occurring a few seconds earlier. Please note that none of the points below chance level was statistically significant and should thus be attributed to random fluctuation.

We also looked for brain areas that encoded the decision after it was made, during the task preparation and execution phase. The task choice could be decoded from the angular gyrus 4 s after the time of the conscious decision (Fig. 2, Right; P < 0.00005 uncorrected, 5-voxel cluster threshold, 64.2% accuracy). Increase in information began as early as 2 s postdecision and reached statistical significance at 4 s postdecision. Taking into account the hemodynamic delay, this means that the angular gyrus probably began to encode the task choice around the time of conscious decision. This likely reflected the preparation followed by actual performance of the arithmetic task, as the angular gyrus has been found to be involved in the retrieval of overlearned arithmetic facts from memory (24, 25). As expected, there was also a global increase in blood oxygen level–dependent (BOLD) signal in the angular gyrus, peaking around 4–6 s after the decision was made (Fig. S2). However, no significant difference between the two arithmetic tasks was found, even when thresholds were relaxed (P < 0.001 uncorrected), indicating again that the choice of task was encoded in the local detailed spatial patterns of activation, rather than the global activation magnitude.

We also used additional searchlight analyses to investigate the neural encoding of motor responses. As expected, the movement could be decoded from bilateral motor cortices and supplementary motor area (SMA), 6–10 s after the abstract decision was made (Fig. S3). Please note that this delay includes the time to complete the calculation task plus the hemodynamic delay. Importantly, these motor regions did not encode the abstract intention (Figs. S2 and S3). In our previous study (2), SMA also had choice-predictive information for motor intentions, albeit later than the parietal and frontopolar cortex. However, it was not possible here to decode the abstract intention from activity patterns in SMA at any time before or after the decision (Fig. S2). In return, the cortical regions that encoded the abstract decisions here did not encode the motor response at any time (Fig. S3), providing further evidence for a dissociation between preparation of mental calculation and motor processing.

As in a previous study (2), we also assessed the prediction of the timing of the decision (as opposed to the specific choice for adding or subtracting). This is referred to as when-versus-what decisions (26, 27). First, we identified brain regions containing early information about the timing of the upcoming decision. Classifiers were trained to distinguish the functional activity between five different points occurring from −8 to 0 s before a decision was made (see Experimental Procedures for full details). This information was found to be encoded in the pre-SMA, extending into the SMA and rostral cingulate zone (P < 0.00005 uncorrected; 5-voxel cluster threshold; 31.4% accuracy; chance level, 20%; Fig. S4), as reported previously for simple motor intentions (2). In contrast, the medial frontopolar and posterior cingulate/precuneus regions that encoded the content of the decision did not contain predictive information about the timing of the decision (t < 1). Next, we explored how accurately the timing of each decision could be predicted before it was consciously made on a trial-by-trial basis. Classifiers were trained to identify a combination of spatial and temporal brain activity patterns occurring in the pre-SMA region from −4 to 0 s before participants made a conscious decision. By detecting when this pattern occurred during each trial, we were able to accurately predict the exact time that participants were going to make a decision before they had made any behavioral response (71.8%; SE = 1.6%; Experimental Procedures).

Finally, we investigated how the predictive signals were related to activity in the task-negative DMN. This was done for two reasons: First, brain regions with choice-predictive information are in close anatomical proximity to regions of the DMN (15–17, 28), thus raising the question of whether they reflect a similar underlying mechanism. Second, high levels of DMN activity are typically observed during off-task periods (17, 28, 29); that is, during in-between periods in which participants are busy performing tasks. This raises the question of whether DMN activity is also high in the period preceding the free decision. Although this would not constitute final evidence, it could nonetheless lend further support to the notion that participants are not busy thinking about the upcoming decision during this period.

The DMN was identified with standard techniques using group-level independent components analyses (ICA) (30), and its average time course was extracted (Fig. 3 and Fig. S5; see Experimental Procedures for details). The same analysis was also applied to data from an earlier experiment involving free motor decisions for comparison (2). In both experiments, a typical frontoparietal DMN profile was observed (Fig. S5). To formally assess the spatial overlap between choice-predictive regions and the DMN, we performed a conjunction analysis (Fig. 3A). In the abstract intention task, there was a partial overlap with DMN in anterior medial prefrontal cortex. For the previously published motor task (2), there was a partial overlap with DMN in medial parietal cortex. Please note that it has been repeatedly shown that successful pattern classification does not depend on whether a brain region shows a net positive (or negative) change in magnitude (2).

Open in a separate window

Fig. 3.

Changes in overall default mode network activity for spontaneous motor and abstract intentions. (A) For both motor (Left) and abstract (Right) tasks, the DMN (blue; P < 0.0001) and choice-predictive (red; P < 0.0001) regions are projected onto the medial cortical surface, with overlapping voxels rendered in green. Please note that volume rendering projects regions from different depths onto the medial plane. (B) Activation time courses of the DMN (black line, left axis) are plotted together with the classification accuracies for precuneus/posterior cingulate (green dotted line, right axis) and frontopolar cortex (cyan dotted line, right axis). For both motor (Left) and abstract (Right) tasks, the default mode activity and the predictive information peaked roughly around the same time before conscious decision. The DMN during this early preparatory phase still resembled typical off-task or “resting” period activity, whereas parietal and prefrontal signals already encoded the upcoming choice.

We also directly compared the time courses of default mode activity and early predictive information and found that the default mode activity and the predictive information peak roughly around the same time (Fig. 3B). The main difference was that the default mode activity began to rise earlier and lasted longer. The comparable peak time of DMN and choice-predictive information is interesting because it means that the DMN exhibits off-task-like behavior, whereas parietal and prefrontal signals already have information about the outcome of the next choice.

Discussion

Our results show that regions of medial frontopolar cortex and posterior cingulate/precuneus encode freely chosen abstract intentions before the decisions have been consciously made. Previously, a similar network of regions was reported to be involved in the unconscious preparation of motor intentions (2). In contrast, in the current experiment, the intended “action” was a nonmotor, abstract mental operation. Importantly, the randomized response mapping ensured that the motor output used to indicate the arithmetic answer was independent of, and did not contain any information about, the abstract decision made. In addition, these regions did not encode the motor response at any time before or after the conscious decision, even when the behavioral response was made. Thus, by fully dissociating high-level intentions from motor preparation and execution, our current findings provided direct evidence that the medial frontopolar and precuneus/posterior cingulate regions were involved in the formation of high-level intentions for voluntary actions in general. This also confirmed that the frontoparietal network previously reported was indeed involved in the formation of free intentions per se (2), rather than motor preparation. In future studies, it should be possible to enhance the sensitivity for smaller differences between informative brain regions by directly comparing motor and calculation tasks in the same participants.

Closer to and after the point of conscious decision, other regions involved in the actual execution of the specific voluntary action began to encode the intention. In the case of a simple motor action, this was the motor system, including SMA and bilateral primary motor cortices (2, 5). For the abstract mental arithmetic task reported here, this was the angular gyrus, involved in the retrieval of overlearned arithmetic facts from memory (24, 25). These areas were only involved in the execution of specific tasks and did not encode intentions in general.

Previous studies have also shown that abstract decisions can be decoded from the medial frontopolar cortex, but only when these decisions were consciously maintained during a delay period before execution (31, 32). Thus, it has remained unclear whether the same regions were also involved in the unconscious formation of voluntary intentions or merely store the decisions after they have been consciously formed. In the current study, participants were not cued to make decisions at specific points in time but were allowed to make decisions spontaneously. By asking participants to report when they first consciously decided, we could investigate what happened in the brain before the decisions were consciously made. We found that both medial frontopolar cortex and posterior cingulate/precuneus started to encode the specific outcome of the abstract decisions even before they entered conscious awareness. Our results suggest that, in addition to the representation of conscious abstract decisions, the medial frontopolar cortex was also involved in the unconscious preparation of abstract decisions.

Both regions that encoded the content of the decision ahead of time have also been implicated in tasks involving both prospective memory (33–37) and imagining both the past and the future (36, 38–41). A possible interpretation of the current results is that these regions were involved in unconscious preparation for actions in the near future (40). It remains unclear whether both conscious and unconscious representations are subserved by the exact same neural substrates within these regions or whether they are separable at a finer scale. Future studies could address this question by verifying whether mutual classification between unconscious and conscious decision representation is possible. If they do share common neural substrates, then it should be possible to use classifiers trained on unconscious decision representation to decode consciously maintained decisions and vice versa.

At present, the specific functional roles of the frontopolar and precuneus/posterior cingulate regions in the formation of free decisions are not fully understood. Interestingly, medial frontopolar and posterior cingulate cortex have also been implicated in other types of decision making involving rewards, rather than “random” and “free” choices (42–47), and in tracking the subjective values of the rewards (48, 49). In most studies of decision making, participants are typically asked to decide between different alternatives that are seldom as balanced as in the current experiment. The choices are often biased by prior feedback (31) or reward characteristics such as reward type, reward level, or reward timing (45, 47, 48).

An important finding was that regions that predicted the content of the decision (adding or subtracting) did not predict the exact timing of the decision. Rather, this information was found in the pre-SMA, which has also been shown to be involved in the generation of self-initiated and self-paced motor actions in experiments in which the timing of decisions was not explicitly measured (1, 2, 5, 50–57). Taken together, when decision time is unconstrained and self-paced, unspecific activity builds up in this region over a few seconds before the decision for a voluntary action, whether motoric or abstract, is triggered. When different alternatives are available, the specific choice is shaped by activity in other regions such as the frontopolar and precuneus/posterior cingulate. Combining both sources of information about intention content and timing may provide a promising approach to understanding the neural mechanisms underlying voluntary actions.

Consistent with our findings, there is increasing evidence that prior fMRI signals can be used to predict certain upcoming mental states: Spontaneous generation of motor intentions (2) and abstract intentions (as here) can be predicted a few seconds before conscious decisions are made, and behavioral errors can be predicted up to 30 s before they occur (58). Although predictive fMRI signals have to be interpreted carefully (59), there is evidence suggesting that they could reflect low-frequency local field potentials (60, 61). Moreover, other studies using electrical recordings have also found early predictive signals. In their original EEG study, Libet et al. (1) reported the onset of the readiness potential to occur a few hundred milliseconds before the conscious motor intention. In a more recent study using single and multiunit depth electrodes, Fried et al. (5) found that in some cases, the firing rates of certain medial frontal neurons could start changing up to a few seconds before the conscious decision is made (see figure S3e in ref. 5), lending support to our earlier (2) and current fMRI findings.

It is interesting that mental calculation, the more complex task, had less predictive lead time than a simple binary motor choice in a previous study (2). This could tentatively reflect a general limitation of unconscious processing in the sense that unconscious processes might be restricted in their ability to develop or stabilize complex representations such as abstract intentions. It is also worth noting that both studies showed the same dissociation between cortical regions that were predictive of the content versus the timing of the decision. This implies that the formation of an intention to act depends on interactions between the choice-predictive and time-predictive regions. The temporal profile of this interaction is likely to determine when the earliest choice-predictive information is available and might differ between tasks.

There was a partial spatial overlap between the choice-predictive brain regions and the DMN. Interestingly, the state of the DMN during the early preparatory phase still resembled that during off-task or “resting” periods. This lends further credit to the notion that the preparatory signals were not a result of conscious engagement with the task (17, 62, 63). Furthermore, the spatial and temporal overlap hints at a potential involvement of the DMN in unconscious choice preparation.

To summarize, we directly investigated the formation of spontaneous abstract intentions and showed that the brain may already start preparing for a voluntary action up to a few seconds before the decision enters into conscious awareness. Importantly, these results cannot be explained by motor preparation or general attentional mechanisms. We found that frontopolar and precuneus/posterior cingulate encoded the content of the upcoming decision, but not the timing. In contrast, the pre-SMA predicted the timing of the decision, but not the content.

Experimental Procedures

Supplementary Material

Acknowledgments

Footnotes

References