Predicting risky choices from brain activity patterns

Abstract

Previous research has implicated a large network of brain regions in the processing of risk during decision making. However, it has not yet been determined if activity in these regions is predictive of choices on future risky decisions. Here, we examined functional MRI data from a large sample of healthy subjects performing a naturalistic risk-taking task and used a classification analysis approach to predict whether individuals would choose risky or safe options on upcoming trials. We were able to predict choice category successfully in 71.8% of cases. Searchlight analysis revealed a network of brain regions where activity patterns were reliably predictive of subsequent risk-taking behavior, including a number of regions known to play a role in control processes. Searchlights with significant predictive accuracy were primarily located in regions more active when preparing to avoid a risk than when preparing to engage in one, suggesting that risk taking may be due, in part, to a failure of the control systems necessary to initiate a safe choice. Additional analyses revealed that subject choice can be successfully predicted with minimal decrements in accuracy using highly condensed data, suggesting that information relevant for risky choice behavior is encoded in coarse global patterns of activation as well as within highly local activation within searchlights.

Keywords: decision-making; fMRI; machine learning.

Fig. 1.

Sample trials from the BART task. On each balloon, subjects made successive choices about whether to pump or cash out. They earned five points for each pump made before cashing out; however, if the balloon exploded before they cashed out, no points were earned for that balloon. Thus, each pump opportunity represented a risky decision where the subject could choose a certain, safe option (cash out) or an uncertain, risky option (pump). Trials boxed in blue indicate trials suitable for classification analysis; trials are matched for level of risk and subject choice on the trial but differ on subject choice on the subsequent trial. Note that although many “pre-pump” trials are present in the figure, only the boxed prepump trial was selected because it has the same level of risk as the paired “pre-cash out” trial.

Fig. 2.

Classification searchlight analysis of pre-cash out vs. prepump trials. Colored voxels indicate centers of searchlights where the classifier could successfully discriminate between prepump and pre-cash out activation patterns (searchlight classification >60%, whole-brain cluster-corrected P < 0.05 via comparison with 1,000 random permutations). (Scale: 60–70%.) Activation maps were projected onto an inflated average cortical surface of the Population-Average, Landmark-, and Surface-Based (PALS) atlas using the multifiducial mapping technique of Van Essen (29).

Fig. 3.

Plot of sample values for the two parameters in the two-parameter classification. The x-axis values indicate mean Z-statistic values for the contrast between prepump and pre-cash out activity in regions where prepump activation is greater than pre-cash out activation (t > 2.0). The y-axis values indicate the same for regions where pre-cash out activation is greater than prepump activation. In each case, voxels were selected using independent training set data. Blue indicates prepump samples, and red indicates pre-cash out samples. The black line reflects the logistic regression classification boundary; this classifier was able to separate prepump and pre-cash out trials successfully (67% success rate).

Fig. 4.

Searchlights that could successfully discriminate between prepump and pre-cash out trials overlaid on a map of prepump vs. pre–cash-out activation thresholded at t > 2.0. More searchlights were located in regions where activation is higher on pre-cash out trials (8,520 voxels of 43,995 total voxels in the univariate mask) than on prepump trials (1,706 voxels of 35,215 total voxels in the univariate mask), suggesting that it is changes in regions related to risk aversion that most reliably predict whether a subject will make a risky or safe choice.