Balancing Caution and Greed: Neurometric Responses to Decision-Making under Escalating Risk

Introduction:
Across a diversity of environments from foraging to financial investment, agents face escalating potential reward and risk (the first being the motivation for accepting the second) and need to retrieve relevant information from the environment in order to update which action to take next.
Methods:
18 healthy volunteers performed a gambling task while task related activity was assessed during a 25-min session of whole-brain fMRI at 3T (Siemens Verio, Erlangen, Germany). During fMRI, subjects played a dice game where they could gain cumulative amounts according to the number of eyes on the die. On each trial, participants could decide to stop to throw the dice and save the accumulated sum, or to continue under the risk of losing the amount in the event of a "1". SPM8 was used to model the neural effects in a general linear model of 3 main conditions of interest: "Continue", "Loss", and "Stop". For "Continue" trials, we modelled parametric modulations of neural activity depending on consecutive outcomes: the accumulated sum gained, the stop-probability (based on logistic choice modelling with sum as predictor) which indicated the subjective evidence of choosing the "safe" option, and the prediction error (actual outcome - expected outcome).
The accumulated sum reflects the objective information on the stake risked by choosing the "Continue" response. The two other parametric modulators are contextual factors and reflect internal parameterizations on the external input of sum: while stop-probability reflects a context of increasing decision-conflict given the current sum, the prediction error signal the current gain given the stake put at risk at the previous trial. A high prediction error reflects a salient event, namely a better-than-expected outcome on top of a high stake, requiring a deliberate choice, and possible a new future action set, rather than an automatic "continue" response as at lower stakes.
Results:
In the following we report significant findings at p<0.05, FWE corrected at cluster level; extent threshold: 0.001.
The accumulated sum was associated with increases in neural activity in a large contiguous cluster with bilateral peaks located in the head of the caudate nucleus extending to ventral striatum (VS), pre-supplementary motor area (pre-SMA), medial prefrontal cortex, bilateral inferior frontal gyri (IFG). "Stop probability" was associated with a parametric modulation of neural activity in dorsal anterior cingulate cortex (dACC), insula, middle frontal gyrus (MFG), and parietal cortices, bilaterally. The "prediction error" was mirrored by neural activity in bilateral primary visual areas, left posterior Insula, dorsal posterior cingulate cortex (dPCC). Additional clusters were seen in left inferior temporal gyrus, left and right superior frontal gyri, and left putamen.
Conclusions:
The above analysis afforded opportunity to distinguish three variables, pertaining to the sequential trade-off of risk and reward produced by consecutive outcomes when repeatedly throwing a die. We identify three partially overlapping sets of brain regions, which scale their activity to these metrics: Activity in bilateral IFG and pre-SMA monitored the accumulated sum, possibly reflecting a build-up of evidence in favor of an alternative "Stop" response (1, 2). Furthermore, striatal regions such as caudate head and VS showed a robust linear scaling to the accumulated sum, which may be caused by a ramping up of tonic dopaminergic activity or phasic releases that scale with the current gain (3, 4). The correlation of dACC and insula activity with stop probability possibly reflects the increasing response conflict (5). The prediction error signal in primary visual areas might indicate a top-down influence on V1, whereas the prediction error signal in dPCC activation fits to theories of its role in balancing between internal and external attentional processing (6).