Neural predictors of risky choice replicate and generalize within and beyond the laboratory

903 

Abstract Submission 

Leili Mortazavi¹, Charlene Wu², Elnaz Ghasemi¹, Brian Knutson³

¹Stanford University, Stanford, CA, ²Toyota Research Institute, Palo Alto, CA, ³Stanford University, Palto Alto, CA

Leili Mortazavi  
Stanford University
Stanford, CA

Charlene Wu, PhD  
Toyota Research Institute
Palo Alto, CA

Elnaz Ghasemi  
Stanford University
Stanford, CA

Brian Knutson, PhD  
Stanford University
Palto Alto, CA

Neuroimaging researchers can currently use neural activity to predict risky choice in humans but consensus on underlying mechanisms has remained elusive (Knutson & Huettel, 2015). Recently, the replicability and generalizability of neuroimaging findings have additionally been questioned (Marek et al., 2022). Thus, identifying and optimizing neural predictors of risky choice requires verifying integrity of both neural and behavioral measures.

Risky gambles minimally require balancing potential positive against negative uncertain outcomes. Dissociation of anticipatory choice processes from sensory input and motor output might identify the most generalizable predictors of risky choice. Thus, we combined a temporally staged task with converging analyses to localize neural predictors of risky choice in humans in three steps. First, we applied Volume-Of-Interest (VOI) and whole-brain analyses to functional MRI data in an original dataset, followed by replication in an independent sample and generalization to two additional samples using a different task. Second, we tested whether identified neural predictors could be dissociated from sensory and motor correlates of risky choice. Third, we investigated whether neural predictors of risky choice could predict individual differences in risk preferences within and beyond the laboratory.

In the original sample (N=75; age: 26±7; 25 female), healthy subjects chose between gambles and a safe option while being scanned with FMRI (optimized for detection of subcortical activity; Srirangarajan et al., 2021). Multivariate prediction of trial-by-trial risky choice was derived from pre-choice activity (2s) in three predicted VOIs (i.e., in the Nucleus Accumbens (NAcc), Anterior Insula (AIns) and Medial PreFrontal Cortex (MPFC)) or across the whole brain using Support Vector Machines with Recursive Feature Elimination (SVM-RFE).

Each analysis was first conducted on the original dataset. After optimizing processing steps and model parameters, identical preprocessing and modeling procedures were applied to raw data from previously published samples using the same task (N=32; age: 52±20; 17 female; Leong et al., 2016) or a different risky choice task (N=15; Tom et al., 2007; N=108; Botvinik-Nezer et al., 2020).

Across four samples (total N=230) and two tasks, NAcc pre-choice activity robustly predicted risk-seeking choices but AIns activity predicted risk-averse choices. These predictive patterns recurred in both VOI-based and whole-brain analyses.

As predicted, analysis of the original sample with SVM-RFE recovered features consistent with a triple dissociation (Figure 1a-b), in which visual cortex activity classified spatial position of the risky gamble, mesolimbic activity classified upcoming risky choice, and motor cortex activity classified laterality of the button press response (Figure 1c).

Across individuals, NAcc activity did not correlate with risk preferences (Rho=–0.11, p=.36) but MPFC activity was associated with risk seeking (Rho=0.39, p=.0008), and AIns activity was associated with risk aversion (Rho=–0.24, p=.034). Further, individual differences in AIns activity during presentation of negatively-skewed gambles was associated with less real-world debt (Rho=–0.39, p=.034).

This work demonstrates not only that neural predictors of risky choice can be dissociated from sensorimotor components, but also that these predictors replicate across samples, generalize to other similar risky choice tasks, and hold some validity for explaining risk preferences both within and beyond the laboratory. Theoretically, the findings help resolve a persistent neuroeconomic question by indicating that two opposing anticipatory affective signals drive risk seeking versus risk avoidance. Practically, the findings also highlight a generalizable set of features for predicting risky choice in humans, and for causally manipulating neural targets in other species.

Emotion and Motivation Other

Decision Making ¹

Activation (eg. BOLD task-fMRI) ²

Multivariate Approaches

Subcortical Structures

Cortex

FUNCTIONAL MRI

Multivariate

Sub-Cortical

Univariate