Pre-acquired Functional Connectivity Predicts Choice Inconsistency

Abstract

Economic choice theories usually assume that humans maximize utility in their choices. However, studies have shown that humans make inconsistent choices, leading to suboptimal behavior, even without context-dependent manipulations. Previous studies showed that activation in value and motor networks are associated with inconsistent choices at the moment of choice. Here, we investigated if the neural predispositions, measured before a choice task, can predict choice inconsistency in a later risky choice task. Using functional connectivity (FC) measures from resting-state functional magnetic resonance imaging (rsfMRI), derived before any choice was made, we aimed to predict subjects' inconsistency levels in a later-performed choice task. We hypothesized that rsfMRI FC measures extracted from value and motor brain areas would predict inconsistency. Forty subjects (21 females) completed a rsfMRI scan before performing a risky choice task. We compared models that were trained on FC that included only hypothesized value and motor regions with models trained on whole-brain FC. We found that both model types significantly predicted inconsistency levels. Moreover, even the whole-brain models relied mostly on FC between value and motor areas. For external validation, we used a neural network pretrained on FC matrices of 37,000 subjects and fine-tuned it on our data and again showed significant predictions. Together, this shows that the tendency for choice inconsistency is predicted by predispositions of the nervous system and that synchrony between the motor and value networks plays a crucial role in this tendency.

Keywords: choice inconsistency; decision-making; functional connectivity; predictive modeling.

Figure 1.

Task and behavioral results. A, Subjects were presented with a budget line with 50–50% lotteries between two amounts, X and Y. Each point on the line represents a different lottery between X and Y, and subjects had to choose their preferred lottery out of all possible lotteries along the budget line. For example, the red dot corresponds to a 50% chance to win 30 tokens (X) and a 50% chance to win 51 tokens (Y). The slopes and endowments of the budget lines were randomized across trials. B, Experimental design. Subjects completed 75 trials divided into three blocks of 25 trials inside the fMRI scanner. C, Subjects’ inconsistency levels measured with Afriat’s index, compared with 1,000 simulated random decision-makers. D, Subjects’ Afriat’s index distribution (log transformed).

Figure 2.

Prediction of inconsistency from resting-state functional connectivity. Correlations between actual and predicted inconsistency levels for Random Forest models based on functional connectivity features derived before the choice task. A, Hypothesized motor–value Random Forest model predictions, using only 10 hypothesized value and motor ROIs ( $r=0.3631$; $p=0.0167$). B, Exploratory whole-brain Random Forest model predictions, using 100 cortical parcellation of the whole-brain ( $r=0.4238$; $p=0.0067$). For results of the Lasso regression models, see Extended Data Figure 2-1.

Figure 3.

Whole-brain Random Forest model’s feature importance. A, Node-wise feature importance color-coded and projected on a brain surface. The most important nodes are the parcels in the left PCC and somatomotor areas. Node $i$’s importance was defined as the average of edge-wise feature importance that include node $i$. B, Top 10 most important edge-wise features. Edge width and shade denote the feature importance. The most important features include the functional connectivity between the PCC, somatomotor areas, and a parcel in the dorsal attention network. Nodes are colored according to the seven resting-state functional networks (Yeo et al., 2011) which are also depicted on the circle’s circumference. For the feature importance of the whole-brain Lasso model, see Extended Data Table 3-1.

Figure 4.

Prediction of inconsistency using the pretrained meta-matching model. A, The meta-matching framework. He et al. (2022) trained a fully connected neural network using the UK Biobank $(n=36,847)$ to predict 67 behavioral and physiological phenotypes from resting-state functional connectivity. Then, a kernel Ridge regression model is trained over the pretrained neural network’s prediction to predict new behavioral traits. We applied this procedure to predict subjects’ inconsistency levels. B, Correlation between actual and predicted inconsistency levels for the meta-matching model ( $r=0.6180$; $p<0.0001$). For the feature importance analysis of the kernel Ridge regression, see Extended Data Table 4-1. For the feature importance analysis of the neural network, see Extended Data Table 4-2.