Anticipating conflict: Neural correlates of a Bayesian belief and its motor consequence

Abstract

Previous studies have examined the neural correlates of proactive control using a variety of behavioral paradigms; however, the neural network relating the control process to its behavioral consequence remains unclear. Here, we applied a dynamic Bayesian model to a large fMRI data set of the stop signal task to address this issue. By estimating the probability of the stop signal – p(Stop) – trial by trial, we showed that higher p(Stop) is associated with prolonged go trial reaction time (RT), indicating proactive control of motor response. In modeling fMRI signals at trial and target onsets, we distinguished activities of proactive control, prediction error, and RT slowing. We showed that the anterior pre-supplementary motor area (pre-SMA) responds specifically to increased stop signal likelihood, and its activity is correlated with activations of the posterior pre-SMA and bilateral anterior insula during prolonged response times. This directional link is also supported by Granger causality analysis. Furthermore, proactive control, prediction error, and time-on-task are each mapped to distinct areas in the medial prefrontal cortex. Together, these findings dissect regional functions of the medial prefrontal cortex in cognitive control and provide system level evidence associating conflict anticipation with its motor consequence.

Introduction

The ability to proactively adjust our behavior is integral to survival. Studying the neural bases of proactive control advances our understanding of how decisions are made in a changing environment and why individuals are engaged in impulsive behavior.

Proactive control has been studied in the laboratory with a variety of behavioral paradigms (Brass and Haggard, 2007; Horga et al., 2011; Kuhn et al., 2009). Frontal and parietal cortices respond to cued attention allocation (Luks et al., 2007) and preparatory control of a switch in response (Rushworth et al., 2001). When participants withheld movements while waiting to detect a target, activation of the superior medial prefrontal cortex (MPFC) and inferior parietal lobule supports proactive control (Jaffard et al., 2008). The importance of proactive control is demonstrated in a computational model of saccadic eye movement (Lo et al., 2009) and may be generalized to other systems (Ballanger, 2009).

In the stop signal task (SST), increased stop signal probability bolsters proactive control, evidenced by delayed activity in the primary motor cortex (Jahfari et al., 2010). Varying the occurrence of go trials prior to a stop trial, Vink et al. (2005) showed MPFC, caudate and left insula increasing activation to stop likelihood. Chikazoe et al. (2009) used an SST with two types of go trials, ‘go-certain’ and ‘go-uncertain’. Motor responses may require interruption in ‘go-uncertain’ trials; thus elicited activations are thought to reflect proactive inhibitory control. A recent study of choice SST showed activation of the superior medial frontal and inferior frontal cortices when participants are informed, compared to uninformed, as to which effector to use (Smittenaar et al., 2013). Furthermore, in reaction time tasks proactive control is frequently followed by prolonged response times, and studies have also described medial prefrontal activities in association with time-on-task (Carp et al., 2010; Grinband et al., 2011). Together, these studies highlighted an important role of the MPFC in proactive control but it remains unclear whether distinct regions of the MPFC mediate conflict anticipation and behavioral outcome or whether these activities are related. Another important issue concerns the confound of stimulus prediction error, which is known to drive MPFC activation (Glascher et al., 2010; Ide et al., 2013; Nee et al., 2011; So and Stuphorn, 2012). As pointed out earlier, because the expectation of the stop signal is not realized during go signal onset, a violation of this expectation or prediction error occurs at the same time, presenting a confound to proactive control (Zandbelt et al., 2013).

The current study aimed to address these issues. We used a Bayesian model to compute the likelihood of stop signal – p(Stop) – trial by trial in the SST and established a correlation between p(Stop) and reaction time (RT) – a sequential effect – for individual subjects. We modeled the fMRI signals at trial onset to characterize activations to p(Stop), and at go signal onset to characterize activations to prediction error and RT slowing. With exclusive masking we identified neural responses specific to each component process of proactive control. We showed that stop signal anticipation, stimulus prediction error, and RT slowing (“time-on-task”) are mapped to distinct areas in the MPFC. Importantly, neural activities specific to stop signal anticipation are both correlated and Granger causally related to activities specific to RT slowing, supporting a directional link between these two processes.

Methods

Participants and behavioral task

One hundred fourteen healthy adults (64 females; 30.7 ± 11.0 years of age) participated in this study. All participants signed a written consent after they were given a detailed explanation of the study in accordance with a protocol approved by the Yale Human Investigation Committee. We were able to include a large number of participants in the current study by combining data sets both from studies exclusively of healthy individuals (Hendrick et al., 2010; Hu et al., 2012; Ide and Li, 2011; Zhang and Li, 2012; Zhang et al., 2012) and from the healthy cohort of clinical studies (Bednarski et al., 2012; Hendrick et al., 2012; Li et al., 2009b; Yan and Li, 2009), all conducted under the same behavioral task and in the identical 3 T scanner.

Participants performed a stop signal task or SST (Hu and Li, 2012; Li et al., 2009a), in which go and stop trials were randomly intermixed in presentation with an inter-trial-interval of 2 seconds (s). A fixation dot appeared on screen to signal the beginning of each trial. After a fore-period varying from 1 s to 5 s (uniform distribution), the dot became a circle – the “go” signal – prompting participants to quickly press a button. The circle disappeared at button press or after 1 s if the participant failed to respond. In approximately one quarter of trials, the circle was followed by a ‘cross’ – the stop signal – prompting participants to withhold button press. The trial terminated at button press or after 1 s if the participant successfully inhibited the response. The time between the go and stop signals, the stop signal delay (SSD), started at 200 ms and varied from one stop trial to the next according to a staircase procedure, increasing and decreasing by 67 ms each after a successful and failed stop trial (Levitt, 1971). The 67 ms step reflects four screen frames (monitor refreshing rate = 60 Hz) during stimulus presentation, as used in almost all of our previous studies (Chao et al., 2012; Farr et al., 2014; Ide et al., 2013; Liao et al., 2014; Luo et al., 2013; Winkler et al., 2013; Zhang et al., 2014, in press). In an earlier work we have tried a smaller staircase (32 ms) in a pilot study, which yielded similar performance profiles (Li et al., 2005). With the staircase procedure we anticipated that participants would succeed in withholding the response half of the time. Participants were trained briefly on the task before imaging to ensure that they understood the task. They were instructed to quickly press the button when they saw the go signal while keeping in mind that a stop signal might come up in some trials. In the scanner, they completed four 10-minute sessions of the task, with approximately 100 trials in each session.

Behavioral data analysis

A critical SSD was computed for each participant that represented the time delay required for the participant to successfully withhold the response in half of the stop trials, following a maximum likelihood procedure (Wetheril et al., 1966). Briefly, SSDs across trials were grouped into runs, with each run being defined as a monotonically increasing or decreasing series. We derived a mid-run estimate by taking the middle SSD (or average of the two middle SSDs when there was an even number of SSDs) of every second run. The critical SSD was computed by taking the mean of all mid-run SSDs. It was reported that, except for experiments with a small number of trials (<30), the mid-run measure was close to the maximum likelihood estimate of X50 (50% positive response, Wetheril et al., 1966). The stop signal reaction time (SSRT) was computed for each participant by subtracting the critical SSD from the median go trial reaction time (Logan et al., 1984).

The SSRT can also be computed from a critical SSD estimated from an “inhibitory function”, by fitting the response rates at different SSDs to a sigmoid function. In order to have a robust estimate of correct response rates, the trial numbers at different SSDs should be the same or similar, which applies to experimental designs where the SSD’s are blocked. In the current work, we used a staircase procedure with varying trial numbers across SSDs. In particular, the low number of trials at the lower and higher ends of SSD oftentimes resulted in a less than ideal fit. Therefore, we used a maximum likelihood procedure to estimate the critical SSD and SSRT, following many of our published studies and the literature.

A sequential effect was quantified by Pearson correlation between p(Stop) – the Bayesian estimate of the probability of a stop signal (see below) – and RT on go trials for individual subjects.

Trial by trial Bayesian estimate of the likelihood of a stop signal

We used a dynamic Bayesian model (Yu and Cohen, 2009) to estimate the prior belief of an impending stop signal on each trial, based on prior stimulus history. The model assumes that subjects believe that stop signal frequency r_k on trial k has probability α of being the same as r_{k – 1}, and probability (1 – α) of being re-sampled from a prior distribution π(r_k). Subjects are also assumed to believe that trial k has probability r_k of being a stop trial, and probability 1 – r_k of being a go trial. Based on these generative assumptions, subjects are assumed to use Bayesian inference to update their prior belief of seeing a stop signal on trial k, p(r_k|s_{k – 1}) based on the prior on the last trial p(r_{k – 1}|s_{k – 1}) and last trial’s true category (s_k = 1 for stop trial, s_k = 0 for go trial), where s_k = {s₁,…,s_k} is short-hand for all trials 1 through k. Specifically, given that the posterior distribution was p(r_{k – 1}|s_{k – 1}) on trial k – 1, the prior distribution of stop signal in trial k is given by:

$$p(r_k\mid s_{k-1})=\alpha p(r_{k-1}\mid s_{k-1})+(1-\alpha )\pi \left(r_k\right),$$

where the prior distribution π(r_k) is assumed to be a beta distribution with prior mean pm, and shape parameter scale, and the posterior distribution is computed from the prior distribution and the outcome according to Bayes’ rule:

$$p(r_k\mid s_k)\propto P(s_k\mid r_k)p(r_k\mid s_{k-1}).$$

The Bayesian estimate of the probability of trial k being a stop trial, which we colloquially call p(Stop) in this paper, given the predictive distribution p(r_k|s_{k – 1}) is expressed by:

$$\begin{array}{cc}\hfill P(s_k=1\mid s_{k-1})=& \int P(s_k=1\mid r_k)P(r_k\mid s_{k-1})d{r}_k\hfill \\ \hfill =& \int r_{k}P(r_k\mid s_{k-1})d{r}_k=\langle r_k\mid s_{k-1}\rangle .\hfill \end{array}$$

In other words, the probability p(Stop) of a trial k being a stop trial is simply the mean of the predictive distribution p(r_k|s_{k – 1}). The assumption that the predictive distribution is a mixture of the previous posterior distributions and a generic prior distribution is essentially equivalent to using a causal, exponential, linear filter to estimate the current rate of stop trials (Yu et al., 2009). In summary, for each subject, given a sequence of observed go/stop trials, and the three model parameters {α, pm, scale}, we estimated p(Stop) for each trial.

Imaging protocol and spatial preprocessing of brain images

Conventional T1-weighted spin-echo sagittal anatomical images were acquired for slice localization using a 3 T scanner (Siemens Trio). Anatomical images of the functional slice locations were obtained with spin-echo imaging in the axial plan parallel to the anterior commissure-posterior commissure (AC–PC) line with TR = 300 ms, TE = 2.5 ms, bandwidth = 300 Hz/pixel, flip angle = 60°, field of view = 220 × 220 mm, matrix = 256 × 256, 32 slices with slice thickness = 4 mm and no gap. A single high-resolution T1-weighted gradient-echo scan was obtained. One hundred seventy-six slices parallel to the AC–PC line covering the whole brain were acquired with TR = 2530 ms, TE = 3.66 ms, bandwidth = 181 Hz/pixel, flip angle = 7°, field of view = 256 × 256 mm, matrix = 256 × 256, 1 mm³ isotropic voxels. Functional blood oxygenation level dependent (BOLD) signals were then acquired with a single-shot gradient-echo echo-planar imaging (EPI) sequence. Thirty-two axial slices parallel to the AC–PC line covering the whole brain were acquired with TR = 2000 ms, TE = 25 ms, bandwidth = 2004 Hz/pixel, flip angle = 85°, field of view = 220 × 220 mm, matrix = 64 × 64, 32 slices with slice thickness = 4 mm and no gap. There were three hundred images in each session for a total of 4 sessions.

Data were analyzed with Statistical Parametric Mapping (SPM8, Wellcome Department of Imaging Neuroscience, University College London, U.K.). In the pre-processing of BOLD data, images of each participant were realigned (motion-corrected) and corrected for slice timing. A mean functional image volume was constructed for each participant for each run from the realigned image volumes. These mean images were co-registered with the high resolution structural image and then segmented for normalization to an MNI (Montreal Neurological Institute) EPI template with affine registration followed by nonlinear transformation (Ashburner and Friston, 1999; Friston et al., 1995). Finally, images were smoothed with a Gaussian kernel of 8 mm at full width at half maximum. Images from the first five TRs at the beginning of each trial were discarded to enable the signal to achieve steady-state equilibrium between radio frequency pulsing and relaxation.

Generalized linear models

Our goal is to identify the correlates of stop signal anticipation – the Bayesian belief of a stop signal – as well as the correlates of absolute/unsigned stimulus prediction error and RT slowing. To this end, we distinguished four trial outcomes: go success (GS), go error (GE), stop success (SS), and stop error (SE) for two generalized linear models (GLM). Because p(Stop) is updated on a trial by trial basis, we posited that activities related to stop signal anticipation should arise at trial onset. Because the onsets of fixation point and go signal were on average 3 s apart, and the canonical hemodynamic response peaks at 6–10 s, it was not feasible to include both events in a single model to identify activations specific to each event (Huettel and McCarthy, 2000, 2001; Soon et al., 2003). It is suggested that an average lag of 6 s between two successive events is required to allow near-full separation (Huettel et al., 2009). Therefore, in the current study, we constructed two separate models, each describing the events of interest with fixation point (F model) and go signal (G model) onsets (Hu and Li, 2012). In the first GLM, the F (for fixation point at trial onset) model, we modeled BOLD signals by convolving the onsets of the fixation point – the beginning – of each trial with a canonical hemodynamic response function (HRF) and the temporal derivative of the canonical HRF (Friston et al., 1995). Realignment parameters in all 6 dimensions were entered in the model. We included the following variables as parametric modulators in the model: p(Stop) of GS trials, SSD of SS trials, p(Stop) of SS trials, SSD of SE trials, p(Stop) of SE trials, in that order. Inclusion of these variables as parametric modulators improves model fit (Buchel et al., 1998; Buchel et al., 1996; Cohen, 1997; Hu et al., 2015). Specifically, the parametric modulator of p(Stop) allowed us to examine the neural correlates of stop signal anticipation. Serial autocorrelation of the time series was corrected by a first degree autoregressive or AR(1) model (Della-Maggiore et al., 2002; Friston et al., 2000). The data were high-pass filtered (1/128 Hz cutoff) to remove low-frequency signal drifts. In the first level analysis, we obtained for each participant a contrast “1” on the parametric modulators “p(Stop)” weighted by the proportion of trial number each of GS, SS, and SE trials to examine how deviations from the average BOLD amplitude are modulated by trial-by-trial estimate of the likelihood of a stop signal (St Jacques et al., 2011; Wilson et al., 2009). That is, this contrast identified voxels with activation increasing with the likelihood that a stop signal would appear.

In the second GLM, the G (for go signal onset) model, we modeled the BOLD signals by convolving the onsets of the go signal of each trial with a canonical HRF and its temporal derivative. The goal is to identify regional activations related to RT while controlling for absolute or unsigned stimulus prediction error (UPE): |stimulus p(Stop)|, where stimulus is 1 for a stop and 0 for a go trial (Ide et al., 2013). Thus, we included the following variables as parametric modulators: |0–p(Stop)| or p(Stop) of GS trials, RT of GS trials, SSD of SS trials, |1–p(Stop)| of SS trials, SSD of SE trials, |1–p(Stop)| of SE trials, and RT of SE trials, in that order. We used a contrast “1” on the parametric modulators of UPE, weighted by the number of GS, SS, and SE trials, to examine activations to UPE, and a contrast “1” on the parametric modulator of go RT to identify activations to increasing go trial RT. Note that, in SPM, a parametric modulator is orthogonalized with respect to its preceding parametric modulator. Thus, by placing UPE before RT we effectively regressed out activities related to UPE in the contrast on go RT.

In the second level analysis, one-sample t tests described the group effect at a threshold of p < 0.05, corrected for family-wise error (FWE) of multiple comparisons, for each contrast. Although BOLD signals were modeled each for fixation and go signal onsets in the F and G models, the temporal proximity of these onsets along with the slow hemodynamic response remains a potential issue to disambiguating correlates of p(Stop) and UPE. Thus, we used exclusive masking (Li et al., 2007; Pochon et al., 2002) to identify activations specific to each contrast, with a conjunction of “F_p(Stop) and ~G_UPE” for stop signal anticipation, “G_UPE and ~F_p(Stop)” for unsigned stimulus prediction error, and “G_RT and ~F_p(Stop)” for RT slowing (F and G each indicating the first and second GLM). Importantly, in order to remove confounding activities, we set the masking contrasts at a liberal threshold: p < 0.005 uncorrected.

Single trial amplitude (STA)

To examine the relationship of the BOLD signals of stop signal anticipation and RT slowing, we computed and cross-correlated the single trial amplitude (STA) of all regions of interest (ROIs) for each individual subject. STAs were obtained by fitting a design matrix containing separate predictors for the onset times of each trial (Eichele et al., 2008). The onsets were convolved with a canonical HRF and with the temporal derivative of the canonical HRF, as in the GLM. Least squares were used to estimate the scaling coefficients (β) in the multiple linear regression model and these β estimates were taken as the STA. Thus, for each individual participant, we extracted the STA at fixation onset of GS trials for the ROIs responding to stop signal anticipation (F model), and the STA at go signal onset of GS trials for the ROIs responding to prolonged RT (G model). These two sets of STA’s were correlated with a Pearson regression for each individual subject. We then used a binomial test to examine which pair-wise correlations are significant across participants. Note that because the BOLD signals of the ROIs associated with stop signal anticipation were modeled at a time point preceding those associated with RT slowing, a significant correlation would suggest a directional link between the two processes.

Granger causality analysis (GCA)

As stop signal anticipation takes place prior to RT slowing, we hypothesized that neural activities associated with stop signal anticipation Granger causes activities associated with RT slowing. To confirm this hypothesis, we employed a multivariate GCA to examine the direction of influence between the ROIs (Deshpande et al., 2008, 2009; Duann et al., 2009; Granger, 1969; Ide and Li, 2011; Stilla et al., 2007).

The multivariate GCA was performed for individual participants. For each subject and each ROI, a summary time series was computed by averaging across voxels inside the ROI. These average time series were concatenated across sessions, after detrending and normalization (Ding et al., 2000). The pre-processed time series were used for multivariate GCA modeling. We used Akaike Information Criterion (AIC), which imposes a complexity penalty on the number of parameters and avoids over-fitting of the data (Akaike, 1974). The multivariate GCA required that each ROI time series was covariance stationary, which we confirmed with the Augmented Dickey Fuller (ADF) test (Hamilton, 1994). The ADF test verified that there was no unit root in the modeled time series. The residuals were used to compute the Granger causality measures (F values) of each possible connection between ROIs. Alternatively, connectivity strength could be measured by using the variance of the residual other than the sum of square of the variable (Geweke, 1982; Goebel et al., 2003), which we referred to as the Geweke test. However, since multivariate GCA modeling often involves highly interdependent residuals (Deshpande et al., 2009), we used permutation resampling (Hesterberg et al., 2005; Seth, 2010) to obtain an empirical null distribution of no causality, as suggested in Roebroeck et al. (2005), in order to estimate the F_critical, and assess the statistical significance of Granger causality measures. With resampling, we produced surrogate data by randomly generating time series with the same mean, variance, autocorrelation function, and spectrum as the original data (Theiler et al., 1992), as implemented in previous EEG (Kaminski et al., 2001; Kus et al., 2004), and fMRI studies (Deshpande et al., 2009). We used both the Geweke test (Geweke, 1982; Goebel et al., 2003) and binomial test to assess statistical significance in group analysis (Duann et al., 2009; Uddin et al., 2014). In the Geweke test, we reported the connections with F value > F_critical estimated by using permutation resampling approach (Seth, 2010). In the binomial test, for each connection, we counted the number of subjects that showed F > F_critical (i.e., significant connection) and estimated its significance using a binomial distribution with parameters n = 112 (the time series of 2 subjects were not covariance stationary) and p = q = 0.5 (same probability to observe a connection or not). Multiple comparisons were corrected for false discovery rate (FDR; Genovese et al., 2002).

Results

Behavioral performance

Across all participants, the median go trial reaction time and stop error trial reaction time were 602 ± 10 ms (mean ± standard error) and 533 ± 9 ms, respectively. Participants responded in 98.6 ± 0.2% of go trials and 51.2 ± 0.3% of stop trials. The average stop signal reaction time (SSRT) was 194 ± 3 ms. These measures are typical of stop signal task performance and suggest that participants’ performance was well tracked by the staircase procedure. Furthermore, across all participants, go trial RT was significantly correlated with p(Stop) – an effect that we term “sequential effect” (r = 0.9665, p = .0000, Figs. 1a and b) – stop error rate is negatively correlated with p(Stop) in a Pearson regression (r = −0.9080, p = .0000, Fig. 1c), as also predicted by the Bayesian model (Ide et al., 2013; Shenoy and Yu, 2011). The sequential effect did not correlate with SSRT (p = 0.190).

Fig. 1.

Bayesian model prediction of behavioral performance in the stop signal task. (a) Correlation between p(Stop) and go trial RT across all go success trials. Gray lines are the fitted regressions for individual participants; black solid and dashed lines are the mean and 95% confidence interval of the regressions. (b) Positive correlation between go trial RT and p(Stop) on each trial collapsed over all participants. Gray dots and bars are the mean and standard error of go trial RT averaged across trials for each p(Stop) bin; black line is the best linear fit to the mean; histogram shows the distribution of p(Stop) bin at intervals of 0.01. (c) Negative correlation between stop error (SE) rate and p(Stop) with the same format as in (b).

Regional activations modulated by p(Stop), UPE, and go trial RT

To control for type I errors, we evaluated all imaging results at p = 0.05, corrected for family wise error (FWE) of multiple comparisons. Anticipation of the stop signal engaged activations in bilateral IPL and anterior pre-supplementary motor area (pre-SMA), right OFC and SFG, cerebellum, and insula. UPE was associated with activations in bilateral IPL and angular gyri, bilateral insula including the IFG, thalamus, pallidum, putamen, and DLPFC, right superior temporal gyrus (STG)/middle temporal gyrus (MTG), left cerebellum, and dorsal anterior cingulate cortex (ACC). Prolonged go trial RT was associated with activation in bilateral insula extending to the IFG pars opercularis and posterior pre-SMA (Table 1; Figs. 2(a)–(c)). As described in detail in the Methods section, we applied exclusive masking in order to identify regional activations specific to each contrast. Among these regions the pre-SMA responded to stop signal anticipation; UPE is associated with activations in bilateral insula and IFG, right thalamus, bilateral pallidum and putamen, bilateral IPL and angular gyri, bilateral DLPFC, left OFC, right SFG, left cerebellum, and dorsal ACC; RT slowing is associated with activations in bilateral anterior insula and posterior pre-SMA (Fig. 2d). We overlaid these activities on a mid-sagittal section and showed that distinct loci in the medial prefrontal cortex respond to stop signal anticipation, UPE, and RT slowing.

Table 1.

Regional activations to stop signal anticipation (F model), unsigned prediction error (G model), and positive RT modulation (G model).

Contrast	Region	Cluster size	Cluster level	Peak voxel	MNI coordinate (mm)
			FWE p value	Z value	X	Y	Z
Stop signal  anticipation	R IPL	738	0.000	6.78	57	−43	49
	R OFC/SFG	463	0.000	5.99	33	56	−5
	L IPL	515	0.000	5.91	−57	−46	40
	L CBL	158	0.006	5.82	−24	−70	−32
	R precuneus	70	0.002^	5.18	9	−70	49
	R/L MCC	72	0.005^	4.99	3	−25	28
	R/L ant. preSMA	181	0.003	4.92	3	35	52
	R insula	74	0.031^	4.56	51	17	1
Unsigned  prediction  error	R IPL/AG	1043	0.000	6.95	54	−40	43
	R insula	1622	0.000	6.88	39	17	−2
	R IFG			5.48	39	50	4
	R DLPFC			5.11	48	11	40
	R thalamus			4.72	15	−13	10
	R pallidum			4.23	24	−1	−2
	L IPL/AG	883	0.000	6.54	−57	−40	49
	R STG/MTG	34	0.000^	5.73	51	−25	−8
	L insula	786	0.000	5.71	−36	17	−2
	L thalamus			4.24	−18	−19	7
	L IFG			4.09	−51	8	10
	L OFC	179	0.002	4.84	−36	41	19
	L CBL	232	0.001	4.74	−39	−55	−26
	R STG/MTG	20	0.020^	4.68	48	−31	−8
	dACC	122	0.013	4.47	6	29	31
Prolonged RT	R insula/IFG po	232	0.000	7.92	33	23	4
	L insula	140	0.007	6.93	−33	20	7
	R post. pre-SMA	70	0.015^	5.01	12	5	61

Note:

^{^}peak voxel p value;

L: left; R: right. IPL: inferior parietal lobule; OFC: orbitofrontal cortex; SFG: superior frontal gyrus; CBL: cerebellum; MCC: middle cingulate cortex; ant. pre-SMA: anterior pre-supplementarymotor area; AG: angular gyrus; IFG: inferior frontal gyrus; DLPFC: dorsolateral prefrontal cortex; STG: superior temporal gyrus; MTG: middle temporal gyrus; dACC: dorsal anterior cingulate cortex; IFG po; inferior frontal gyrus, pars opercularis; post. pre-SMA: posterior pre-supplementary motor area.

Fig. 2.

Regional activations to (a) stop signal anticipation, (b) unsigned prediction error, (c) prolonged reaction time (RT), and (d) distinctive medial frontal cortical activations specific to stop signal anticipation (red), unsigned prediction error (blue), and go trial RT modulation (green) overlaid on a structural template in axial sections (upper row, from z = −20 to 60) and sagittal sections (lower row, from x = 2 to 14).

Relationship between activations to anticipation of the stop signal and to go trial reaction time

The effect sizes of these regional activities are presented in Table 1. We isolated regional activations specific to stop signal anticipation and go RT modulation by exclusive masking, which included anterior pre-SMA to conflict anticipation and posterior pre-SMA and bilateral insula to RT slowing. We compared the normalized contrast value of each of these regions of interest to contrast its role in stop signal anticipation and go RT modulation (Fig. 3). The results showed that these contrasts are significantly different for all four ROIs: anterior pre-SMA (t₍₂₂₆₎ = 5.1915, p = 0.0000); left insula (t₍₂₂₆₎ = 2.1361, p = 0.0337); right insula (t₍₂₂₆₎ = 4.0480, p = 0.0000); and posterior pre-SMA (t₍₂₂₆₎ = 2.0585, p = 0.0407), with anterior pre-SMA showing higher activity during stop signal anticipation while posterior pre-SMA and bilateral insula showing higher activity during RT slowing. These results did not provide new information but simply confirmed the contrast specificity of these regional activations.

Fig. 3.

Standardized contrast values of anterior pre-SMA, bilateral insula, and posterior pre-SMA in (a) F model (stop signal anticipation) and (b) G model (Go trial RT slowing).

Because anticipation and RT modulation were each aligned to fixation and go-signal onset, anticipation-related activations temporally precede RT-related activations. We reasoned that these activities should be correlated to reflect the sequential effect — correlation between p(Stop) and RT. To test this hypothesis, we extracted single trial amplitude (STA) of these brain regions for cross correlation for each individual participant, and used a binomial test to evaluate the significance of the pair-wise correlation for the whole group (Table 2). Across all participants, anticipation-related activity of anterior pre-SMA is positively correlated to activations of the posterior pre-SMA and bilateral insula in RT modulation. These results suggest a directional link between signal anticipation and RT slowing.

Table 2.

The number of participants (out of 112)who showed a correlation in single trial amplitude between regions of interest.

		RT modulation
		R insula/IFG po	L insula	R post. pre-SMA
Stop signal  anticipation	L cerebellum	62	59	71*
	R precuneus	65	57	66
	Ant. pre-SMA	74**	75**	74**

Note:

^{^}The cluster in the white matter of the orbitofrontal cortex is not included;

RT: reaction time; R: right; L: left; IFG po: inferior frontal gyrus, pars opercularis; ant. pre-SMA: anterior pre-supplementary motor area; post. pre-SMA: posterior pre-supplementary motor area.

Binomial test:

^*p < 0.01;

^**p < 0.001.

Although Granger causality analysis (GCA) has its limitations in addressing causal relationship between regional time series (Friston, 2009) and cannot elucidate causal interaction in an event-related manner, it has been widely used to support directional functional connectivity between regional activities (Roebroeck et al., 2005; Seth, 2010). Thus, we used GCA to further confirm the directional connectivity between brain regions responding specifically to stop signal anticipation and prolonged RT, positing that activities reflecting stop signal anticipation should Granger cause activities during RT slowing. Of all 114 participants, 2 did not show covariance stationary time series and were excluded from the analyses. We evaluated the significance of each connection across all four brain regions for individual subjects (p < 0.05, corrected for FDR, Genovese et al., 2002) and performed both the Geweke test and binomial test in group analysis (n = 112). The results of the Geweke test showed a unilateral connectivity from the anterior pre-SMA and all regions of RT slowing. In addition, the posterior pre-SMA showed a unilateral connectivity to bilateral insula (Fig. 4a). The results of the binomial test were largely similar, with the exception of bilateral connectivity between the anterior pre-SMA and posterior pre-SMA (Fig. 4b). The average optimal model order computed with AIC was 2.38; the average (standard deviation) adjusted sum-squared-error was 0.38 (±0.13), which was above the minimum of 0.3 as suggested by Seth (2010).

Fig. 4.

Results of Granger Causality Analysis of the time series of the anterior pre-SMA (stop signal anticipation) and the three ROIs, including the right insula/IFG po, left insula, and posterior pre-SMA (RT modulation) in (a) Geweke test and (b) binomial test with numbers indicating the number of participants with a significant connection. Numbers in parenthesis indicate the p value for that connection.

Discussion

Neural correlates of proactive control — stop signal anticipation and prolonged RT

By estimating the probability of a stop signal or p(Stop) trial by trial with a dynamic Bayesian model, modeling fMRI signals both at trial and signal onsets, and masking exclusively between contrasts, we identified regional activities specific to stop signal anticipation, unsigned stimulus prediction error, and prolonged RT. The anterior pre-SMA responds exclusively to p(Stop), suggesting that this area plays a critical role in proactive control. A number of frontal and parietal as well as subcortical regions and the cerebellum respond to prediction error. The posterior pre-SMA and bilateral anterior insula respond to RT slowing. Moreover, anterior pre-SMA activation temporally precedes and correlates with activations of the posterior pre-SMA and bilateral anterior insula. This directional connectivity is also supported by Granger causality analysis. Together, these findings elucidate neural processes with a directional link between conflict anticipation and RT slowing.

Zandbelt et al. (2013) used a cue to inform participants of stop signal percentage before go signal presentation and reported higher activities in the pre-SMA, bilateral insula, and midbrain during the cue than stimulus presentation, underscoring the importance of these areas in proactive inhibitory control. Here, we specifically showed that anterior pre-SMA activity is associated with trial-by-trial estimate of the stop signal likelihood on the basis of a Bayesian belief model. The anterior pre-SMA is at a location where intentional inhibition occurs and negative motor phenomena were observed with electrical microstimulation (Filevich et al., 2012). Specifically, in a study in which participants initiated an intention to act and then decided to complete or cancel the action, greater activation was observed for a dorsomedial prefrontal region (x = −2, y = 41, z = 32) and bilateral anterior insula during action inhibition (Brass and Haggard, 2007). Although the coordinates are different, this prefrontal cluster and the current locus of anterior pre-SMA are both in the Brodmann area 9 (BA 9), suggesting a distinct role of this prefrontal region in proactive control. Future studies can determine whether the anterior pre-SMA mediates a delayed, controlled action (as in the SST) and the more anterior locale of BA 9 mediates a decision to terminate an intended action. Notably, because pre-SMA activity anticipates the motor outcome, this finding is in accord with the proposition that pre-SMA activity is driven by the engagement of attentional or working memory resources and not by stimulus-driven processes typical of reactive control (Criaud and Boulinguez, 2013). More broadly, the finding is consistent with numerous studies supporting a role of pre-SMA in response selection (Krieghoff et al., 2009; Mueller et al., 2007; Rae et al., 2014) and awareness of voluntary action (Ritterband-Rosenbaum et al., 2014; Wolpe et al., 2014).

In rodents, inactivation of the dorsomedial prefrontal cortex impairs the ability to wait for a stimulus, resulting in premature responding (Narayanan et al., 2006), and eliminates phase locking of oscillatory activity with the motor cortex, resulting in loss of behavioral adjustment following errors (Narayanan et al., 2013). In monkeys, neurons in the pre-SMA respond to a switch from automatic to controlled action during a cued switch task (Isoda and Hikosaka, 2007). These no-go neuronal activities are obtained for contralateral, ipsilateral, as well as bilateral eye movements, suggesting that they are upstream from a motor mechanism. Furthermore, microstimulation of the pre-SMA replaces automatic incorrect responses with correct slower responses. In a stop signal task, pre-SMA neurons respond early enough to influence the cancelation of movement (Scangos and Stuphorn, 2010) and encode information about expected and actual outcome, supporting a performance monitoring function (Scangos et al., 2013). Although one needs to be cautioned in cross referencing between species, these and other studies together substantiate the role of pre-SMA in proactive control and highlight the importance in distinguishing executive and stimulus-driven processes of response inhibition (Schall and Godlove, 2012).

Go trial RT prolongs when participants expect to encounter a stop signal. Prolonged go trial RT is associated with activation in bilateral insula and the posterior pre-SMA. Insula activation has been associated with a prolonged RT in auditory syllable identification (Binder et al., 2004) and fear–disgust two-choice discrimination (Thielscher and Pessoa, 2007). We recently simulated stop signal task performance with a leaky accumulator model and showed that both posterior pre-SMA/SMA (x = −6, y = −1, z = 55) and insula increased activation to prolonged RT, manifesting as a slower growth rate in information accumulation (Hu et al., 2014). In a recent study of motion perception, insula activation is associated with a lower percentage of coherence in moving direction of random dots and prolonged response time, supporting an underlying mechanism of slower information accumulation (Ballanger, 2009). On the other hand, the anterior insula is part of the salience network and has been shown to respond to conflicts, errors and prediction errors (Bartra et al., 2013; Eckert et al., 2009; Ham et al., 2013; Harsay et al., 2012; Limongi et al., 2013; Lutz et al., 2013; Manoliu et al., 2014; Nelson et al., 2010; Preuschoff et al., 2008; Rothkirch et al., 2013; Swick et al., 2011; Veldhuizen et al., 2011). Here, we demonstrated that, after accounting for activations related to conflict and error (stop trials as main regressors) as well as prediction error (as parametric modulator) – all of which are highly salient – bilateral anterior insula respond to prolonged RT. This finding supports a multi-faceted role of the anterior insula in cognitive and affective processes (Nelson et al., 2010; Nieuwenhuys, 2012; Uddin et al., 2014).

Posterior pre-SMA is known to be involved in the planning and execution of movements (Boecker et al., 2008; Hoshi and Tanji, 2004; Nakajima et al., 2009; Rowe et al., 2010; Tremblay and Gracco, 2010; Yoshida et al., 2013). In a recent study, Grinband et al. (2011) reported greater activities in the posterior pre-SMA and SMA for prolonged RT, after controlling for conflict in the Stroop task. These regions also showed higher activations for congruent trials with long RT than for in-congruent trials with short RT. Similarly, the posterior pre-SMA was associated with prolonged RT in a multi-source interference task (Carp et al., 2010). Other studies suggest a more complex relationship between posterior pre-SMA activity and RT. For instance, in a rapid serial target detection task to measure sustained attention, participants slowed down in detecting a target toward the later part of the task (Hilti et al., 2013). Compared to poor performers, good performers demonstrated a smaller increase in RT in correlation with sustained activation of the posterior pre-SMA (x = −9, y = 8, z = 43), suggesting that posterior pre-SMA supports rapid target detection. In a study where participants estimated the onset time of intentionally initiated movement, activation of the posterior pre-SMA (x = 2, y = 4, z = 54) is enhanced and correlated with estimate of an earlier than actual onset time (Lau et al., 2006). Taken together, these studies suggest a role of the posterior pre-SMA in cognitive processes that go beyond movement control and the importance of behavioral context in interpreting posterior pre-SMA activation in association with RT.

Distinct loci in the medial prefrontal cortex (MPFC) mediates stop signal anticipation, unsigned stimulus prediction error, and time-on-task

Earlier work has suggested distinct regional activities within the MPFC to mediate action set and prediction error (Rushworth et al., 2004), prediction and outcome evaluation (Jahn et al., 2014), as well as effort allocation and error detection (Oliveira et al., 2014). However, to our knowledge, no studies have aimed to systematically delineate these areal processes within a single behavioral paradigm. Here, the findings highlight that distinct subregions of the MPFC respond to these key processes of cognitive control (Fig. 4): the anterior pre-SMA responds to conflict anticipation, posterior pre-SMA responds to prolonged RT, and the dorsal ACC/SMA responds to unsigned stimulus prediction error.

The functional distinction between anterior and posterior pre-SMA is supported by structural and functional connectivity of the medial frontal cortex. Overall, while the SMA (posterior part of the Brodmann area or BA 6) connects to motor and premotor cortex and responds to movements, the pre-SMA (anterior part of BA 6) connects to other prefrontal structures and responds to cognitive operations whether or not they are related to motor control (Behrens et al., 2006; Johansen-Berg et al., 2004; Klein et al., 2007; Picard and Strick, 1996), a distinction also supported by their connectivity to the basal ganglia and cerebellum (Akkal et al., 2007; Lehericy et al., 2004a; Ramnani et al., 2006). Although less a focus of earlier investigations, the cortex located anterior to pre-SMA (BA 8/9) was distinguished from the pre-SMA in some studies (Lehericy et al., 2004b; Petrides and Pandya, 2007; Zhang et al., 2012). The medial area 8/9 connects to the fronto-polar cortex, dorsolateral prefrontal cortex, posterior pre-SMA, orbitofrontal cortex, posterior cingulate cortex, retrosplenial cortex, dorsal striatum, and dorsomedial thalamus (Petrides and Pandya, 2007). Recently we used resting state fMRI data to parcellate the superior medial PFC on the basis of its whole brain functional connectivity and identified an anteroposterior gradient in connectivity to all of the prefrontal regions and an opposite gradient in connectivity to the primary and premotor cortex (Zhang et al., 2012). In addition, the anterior and posterior pre-SMA each connects to the associative and motor subregions of the thalamus, striatum, and pallidum. These connectivities support a functional distinction between anterior and posterior pre-SMA, in accord with their role each in conflict anticipation and a downstream mechanism that prolongs response time.

Consideration of the current findings in the context of free energy principle (Friston, 2010)

In the current work, we have distinguished regional activations in the MPFC specific to the component processes of cognitive control — conflict anticipation, prediction error and RT slowing. One could consider these results from a free energy principle (Friston, 2010; Friston and Kiebel, 2009) as a general scheme for goal-directed behaviors (Friston, 2012; Friston and Ao, 2012). The free energy principle states that any self-organizing system that is in equilibrium with its environment must minimize its free energy (Friston et al., 2006). That is, a biological entity acts in order to preserve its physical integrity by resisting the natural tendency toward chaos (increases in entropy) according to the second law of thermodynamics. Put into the perspective of a Bayesian decision theory, thus, optimal decisions are those that minimize the free energy (as a cost function). In Bayesian modeling of the SST, we assumed the anticipation of the stop signal – indexed by P(stop) – as the hidden state and the stop/go trials as sensory inputs or observations as well as outcomes. This constitutes our generative model. Across trials, prior beliefs of the model are updated, and posterior probabilities are computed to improve prediction and minimize prediction error. Another way to decrease prediction error would be to make decisions that diminish uncertainty about the outcomes (Friston and Kiebel, 2009). Participants optimize motor responses to minimize the free energy, for instance, by slowing down the response in order to stop successfully when a stop signal appears (Fig. 1c). On the other hand, in the current Bayesian model, we defined outcome as the event – go and stop – which appeared randomly at a pre-determined frequency (~1/4 stop trials). Therefore, an alternative Bayesian model where stop success and error trials distinguish the outcome would be required to examine the role of motor action in minimizing error. An additional consideration is that, in the SST, participants are required to respond within a time window; because this constraint was not made explicit (participants did not receive an error signal when response time exceeded the time limit), we would not be able to fully model the motor decision in the current behavioral task. Together, the free energy principle may provide a useful platform upon which to understand the component neural processes involved in proactive control. However, more work involving alternative models is required to explore the component processes embodied in this conceptual framework.

Conclusions and broader implications

In summary, anterior pre-SMA mediates conflict anticipation prior to activations of the posterior pre-SMA and bilateral anterior insula during RT slowing. These results reveal a cortical circuit of preparatory control, which is essential for flexible human behavior. Because these preparatory processes are often compromised earlier than reactive processes in many neurological conditions (Adams et al., 2012; Cameron et al., 2012; Hakvoort Schwerdtfeger et al., 2012; McLoughlin et al., 2010), they may serve as biomarkers to detect disease onset and to monitor progress of early treatment. For instance, via projection to the striatum, the pre-SMA modulates motor inhibitory control (Rae et al., 2015), a process compromised in Parkinson’s disease (Jahanshahi et al., 2015). Research of cortical striatal interaction in the context of proactive control would advance our understanding of the neural basis of cognitive motor dysfunction in Parkinson’s disease.