Neuro-computational account of how mood fluctuations arise and affect decision making

Abstract

The influence of mood on choices is a well-established but poorly understood phenomenon. Here, we suggest a three-fold neuro-computational account: (1) the integration of positive and negative events over time induce mood fluctuations, (2) which are underpinned by variations in the baseline activities of critical brain valuation regions, (3) which in turn modulate the relative weights assigned to key dimensions of choice options. We validate this model in healthy participants, using feedback in a quiz task to induce mood fluctuations, and a choice task (accepting vs. declining a motor challenge) to reveal their effects. Using fMRI, we demonstrate the pivotal role of the ventromedial prefrontal cortex and anterior insula, in which baseline activities respectively increase and decrease with theoretical mood level and respectively enhance the weighting of potential gains and losses during decision making. The same mechanisms might explain how decisions are biased in mood disorders at longer timescales.

Conflict of interest statement

F.V. has been invited to scientific meetings, consulted and/or served as speaker and received compensation by Lundbeck, Servier and Otsuka. None of these links of interest are related to this work. The remaining authors declare no competing interests.

Figures

Fig. 1

Behavioral tasks. a Trial structure. Each trial included two unrelated tasks. In the quiz task (left), participants had to select one of four possible answers to a general knowledge question. In the choice task (right), participants had to decide between accepting or declining a motor challenge. The two tasks were separated by a rest period. In a subset of participants, subjective ratings of mood (or confidence) were inserted into the rest period. b Motor challenge. The challenge consisted in squeezing a handgrip such that the force peak would fall within a window around 25% of their maximal force. The green line represents a successful trial. Participants could fail because they squeezed too much (upper red line), not enough (lower red line) or for too long (more than 400 ms, black line). The challenge varied across trials along three dimensions that were indicated at the time of choice: difficulty, represented by the size of the target window (in green on a red thermometer), expected gain in case of success, represented by a bunch of regular 10-cent coins (in the upper part) and expected loss in case of failure, represented by a bunch of 10-cent crossed coins (in the bottom part). If subjects declined the challenge, they still had to squeeze the handgrip and try to hit the target window, but the stakes were minimized (expected gain and loss were fixated at 50 cents)

Fig. 2

Computational modeling. The figure illustrates the best models that were selected using group-level random-effect analyses. a Best mood model. Theoretical mood level (TML) was computed as an integration of quiz feedback, with an exponential decay. This feedback was subjective, as its perception was in return biased by TML, and different for positive and negative feedback. TML was then regressed against fMRI data to identify the neural correlates of mood during the rest period between the quiz and choice tasks. b Best choice model. Baseline activity (at the time of prospect display) in mood-related regions (termed Neural Proxy for Mood, NPM) modulated the free parameters of the choice model. A first parameter was the variance of the subjective distribution of forces produced (parameterized by σ, the width of the Gaussian), which served to calculate the subjective probability of success ps, as the integral of probability distribution within the target window. This probability ps was integrated in the expected utility of the prospect, which had two other free parameters—the weights of potential gain and loss, kg and kl. Acceptance probability was calculated as a sigmoid function (softmax) of expected utility, with a last parameter kt that accounted for a linear drift with time

Fig. 3

Choice behavior. The different panels show the effects of experimental factors and neural activity on the probability to accept the motor challenge (in experiment 3, unless indicated otherwise). a Acceptance probability as a function of the prospect dimensions (gain, loss, target size) and the neural proxy for mood (NPM). Squares are binned data points. Dashed line represents acceptance probability as computed by the best choice model. b Left panel: acceptance probability as a function of expected utility. Dashed line represents acceptance probability as computed in the best choice model (without modulation of free parameters by neural activity). Middle panel (experiment 2): Residual error of choice (actual choice – acceptance probability) as a function of mood rating. Right panel: Residual error of choice (actual choice – acceptance probability) as a function of NPM for the best choice model with (dashed line) and without (solid line) modulations of kg by vmPFC baseline and kl by aIns baseline. Note that these modulations captured most of the influence of NPM on the variance in choices that was not explained by task factors. c Acceptance probability as a function of prospect dimensions and NPM, but restricted to trials for which acceptance probability (as computed by the best choice model without modulation of free parameters by neural activity) was between 1/6 and 5/6. Here the effect of baseline activity in mood-related regions was of a similar size as the effects of task factors. Error bars represent inter-subject s.e.m.

Fig. 4

Mood fluctuations. a Variations in mood rating during sessions with positive (green) vs. negative (red) bias, across trials. Bias here means more or less difficult questions, and more or less correct feedback. Note that the seven first and seven last questions of a session (grey windows) were not biased (difficulty was medium). Lines represent means; error bars represent inter-subject s.e.m. The other panels show how well the computational model captured fluctuations in mood level. b Individual differences in model evidence (variational Bayesian approximation to marginal likelihood) between the best mood model and the control model (in which only time was taken into account). Participants were ranked from left to right in ascending order of model evidence. Blue arrow indicates the subject with median model evidence, plotted in the next panel. c Individual example of mood fluctuations across trials. Blue circles are mood ratings (measured or interpolated) and black line is theoretical mood level (TML). See Supplementary Fig. 1 for other individual examples

Fig. 5

Neural activity. The different panels show how baseline activity in key valuation regions reflects mood (top line) and affects computational parameters of the choice model (bottom line). a Statistical parametric map of regions reflecting TML during rest (positive correlation in green, negative correlation in red, cluster generating threshold p < 0.001, cluster size arbitrarily defined for display purpose). Four ROI (vmPFC, aIns, vS, and dmPFC) were defined using spheres of 8-mm diameter centered on coordinates from the literature (coordinates from ref. ). b Impact of ROI activity on choice. Baseline activity was extracted for each subject and trial from our 4 a priori ROI, and orthogonalized with respect to whole-brain mean activity. The 4 trial-by-trial series were then entered in a linear regression model meant to explain the residual error of the choice model fit. Only vmPFC and aIns were significant predictors, with opposite directions. c Contribution of vmPFC and aIns to TML. Baseline activity was extracted at each time point and entered in a linear regression model meant to predict TML. Both vmPFC and aIns were significant predictors of TML, with opposite directions, even when all variables were orthogonalized with respect to the last feedback. d Impact of ROI activity on choice model parameters. Significant modulation was only found with vmPFC for kg (weight on potential gain) and with aIns for kl (weight on potential loss). Bars represent means of parameters estimates; error bars represent inter-subject s.e.m.