Neural correlates of evidence accumulation during value-based decisions revealed via simultaneous EEG-fMRI

Abstract

Current computational accounts posit that, in simple binary choices, humans accumulate evidence in favour of the different alternatives before committing to a decision. Neural correlates of this accumulating activity have been found during perceptual decisions in parietal and prefrontal cortex; however the source of such activity in value-based choices remains unknown. Here we use simultaneous EEG-fMRI and computational modelling to identify EEG signals reflecting an accumulation process and demonstrate that the within- and across-trial variability in these signals explains fMRI responses in posterior-medial frontal cortex. Consistent with its role in integrating the evidence prior to reaching a decision, this region also exhibits task-dependent coupling with the ventromedial prefrontal cortex and the striatum, brain areas known to encode the subjective value of the decision alternatives. These results further endorse the proposition of an evidence accumulation process during value-based decisions in humans and implicate the posterior-medial frontal cortex in this process.

Figure 1. Task design, behavioural and modelling results and EEG.

(a) Schematic representation of the experimental paradigm. After a variable delay (2–4 s), two stimuli (snack items) were presented on the screen for 1.25 s and participants had to indicate their preferred item by pressing a button. The central fixation dimmed briefly when a response was registered. Snack stimuli shown here are for illustration purposes only. Participants viewed real branded items during the experiments. (b) Behavioural performance (red circles) and modelling results (black crosses). Participants' average (N=21) reaction time (RT) and accuracy (top and bottom respectively) improved as the value difference (VD) between the alternatives increased. A sequential sampling model that assumes a noisy moment-by-moment accumulation of the VD signal fit the behavioural data well. (c) Average (N=21) model-predicted evidence accumulation (EA) (black) and EEG activity (red) in the time window leading up to the response (on average, 600–100 ms prior to the response), arising from a centroparietal electrode cluster (darker circles in the inset) that exhibited significant correlation between the two signals (see Methods). Shaded error bars represent standard error across participants.

Figure 2. Average EEG accumulation dynamics as a function of response time, task difficulty and performance.

(a) Participants' average (N=21) response-locked EEG activity from subject-specific best electrodes in the centroparietal cluster in fast and slow trials (defined in terms of the median RT). (b) Participants' average (N=21) response-locked EEG activity from subject-specific best electrodes in the centroparietal cluster in easy and hard trials (defined in terms of the median item value difference). (c) Across subject correlation between the linear slopes of the average response-locked EEG activity over all trials and individual behavioural performance on the task (robust correlation obtained using Wilcox percentage bend correlation, dotted lines: 95% bootstrapped correlations confidence interval).

Figure 3. EEG-informed fMRI and connectivity analyses in the value-based task.

(a) The fMRI GLM model included an EEG-informed regressor capturing the electrophysiological trial-by-trial dynamics of the process of evidence accumulation (EA) in each participant. Three actual single-trial EEG traces are shown. The traces cover the entire trial excluding the time intervals accounting for stimulus processing and motor execution (see inset and Methods for details). To absorb the variance associated with other task-related processes we included three additional regressors: VSTIM—an unmodulated stick function regressor at the onset of the stimuli, VD—a stick function regressor at the onset of stimuli that was parametrically modulated by the value difference between the decision alternatives and RT—a stick function regressor aligned at the time of response and modulated by RT. (b) Hypothetical EA traces in response-locked EEG activity ramping up with different accumulation rates. Convolving these traces with a hemodynamic response function (HRF) leads to higher predicted fMRI activity for longer compared to shorter integration times (that is, the predicted fMRI response scales with the area under each EA trace). (c) The EEG-informed fMRI predictor of the process of EA revealed an activation in pMFC. (d) PPI analysis using the pMFC cluster identified in c as a seed revealed an inverse coupling with a region of the vmPFC and the STR. All activations represent mixed-effects and are rendered on the standard MNI brain at |Z|>2.57, cluster-corrected using a resampling procedure (minimum cluster size, 76 voxels).