Human decisions about when to act originate within a basal forebrain–nigral circuit

Significance Decision-making studies often focus on brain mechanisms for selecting between goals and actions; however, another important, and often neglected, aspect of decision-making in humans concerns whether, at any given point in time, it is worth making any action at all. We showed that a considerable portion of the variance in when voluntary actions are emitted can be explained by a simple model that that takes into account key features of the current environment. By using ultrahigh-field MRI we identified a multilayered circuit in the human brain originating far beyond the medial frontal areas typically linked to human voluntary action starting in the basal forebrain and brain stem, converging in the dopaminergic midbrain, and only then projecting to striatum and cortex.

action times of recent past trials influenced participants' action time on the current trial ( Fig.1D).
Participants were instructed to track the bubbles and to make a response by pressing on a response button at a time of their own choice. Once they responded the stimulus disappeared and a fixation cross appeared on the screen for the interval between response and outcome presentation. The duration of this interval was randomly drawn from a Gamma distribution (min 4 s, max 10 s, mean 5.5 s). During the outcome phase (2 s), if rewarded, a gold, silver, or bronze coin was shown on the screen, representing 20, 10, or 5 p, respectively. If not rewarded, or in rare occasions that participants did not make any response, a dark coin appeared on the screen. In addition to the reward on the current trial, the total reward earned, and the total time left in the experiment was also displayed on the screen at every outcome phase. After the outcome phase the fixation cross reappeared on the screen for the inter-trial-interval (ITI). The duration of the ITI was randomly drawn from a Gaussian distribution (mean = 4.5 s, std = 0.25 s) (Fig.1A).
Before the main task, participants were given written instructions and were trained on the task for 10 min. Once they were comfortable with the task the main experiment started inside the MRI scanner. The duration of each scanning session was set at 45 min. The task finished after 45 min, regardless of the number of trials performed. The experiment was written in Matlab (Mathworks, Natick, USA), using the Psychophysics Toolbox extension (1).
Behavioural analysis. Time to act (actTime) was defined as the natural logarithm of the time passed in seconds from beginning of the trial to the moment that participants made their response. We used a linear mixed-effect model (LMEM) to predict actTime from present and past contextual factors. The maximum likelihood method was used for model estimation. We examined the impact of both present and past contextual factors on actTime.
Present contextual factors consisted of potential reward magnitude, change in reward probability and noise on the current trial. Past contextual factors consisted of actual reward outcome and actTime on the past trial. All predictor variables were normalised. ( (3,4). The model predicts time-to-event (actTime) on the current trial from present and past contextual factors.
Specifically, the predictors (covariates) included reward magnitude, change in reward probability, and ITI of the current trial, and the actual reward and actTime on the past 10 trials. The model is described as: where ( )represents a hazard function (hazard rate of responding), + ( ) represents a baseline hazard function, that is a hazard function when all the covariates are 0, b is a row vector with 23 elements (3 present contextual factors + 10 past rewards + 10 past actTimes) representing Cox coefficients for each covariate and x is a 23 element column vector representing covariates, present contextual factors and contextual factors of the past 10 trials. The coefficients were estimated for each testing session by using the 'coxphfit' function in MATLAB.
A detailed method for obtaining Cox coefficients has been previously described (4). The estimated Cox coefficients ( P ) from the predictors on the current trial and the immediately preceding trial were used to obtain the expected actTime by the following method: First, the cumulative hazard function, Λ R S ( ), of each trial was estimated given the baseline cumulative hazard function, Λ R + ( ), and the covariates: The cumulative hazard function of each trial was then used to estimate the survival function of each trial, S(t): The deterministic actTime is estimated by: Finally, to measure the proportion of variance explained by the Cox regression model, we used Schemper's V (5), which is defined as: Where R is the distance between survival functions of individual trials Si(t) and a survival function estimated from all the trials without taking into account covariates P ( ), by using Kaplan-Meier estimator. R S is calculated in the same way as R , but is the distance between survival functions of individual trials Si(t), and an estimated conditional survival function given covariates x, P S ( ). The equations to calculate R are previously described in detail (4).
Imaging data acquisition. Structural and functional MRI was collected using a Siemens 7 T MRI scanner. High-resolution functional data was acquired using a multiband gradient-echo Whole-brain fMRI data analyses. Whole-brain statistical analyses was performed at twolevels as implemented in FSL FEAT (11). At the first level, we used a univariate general linear model (GLM) framework for each participant to compute the parameter estimates. The contrast of parameter estimates and variance estimates from each scanning session were then combined in a second-level mixed-effects analysis (FLAME 1+2) (12), treating scanning sessions as random effect. The results were cluster-corrected with the voxel inclusion threshold of Z = 3.1 and cluster significance threshold of P = 0.05. The data were prewhitened before analysis to account for temporal autocorrelations (11).
For the first level analysis we looked for brain areas in which activity reflected parametric variation in the empirically observed actTime. Importantly, the GLM only included actTime and not all the present and past contextual factors that influenced actTime.
1: = + + . + 2 + :   and HB activity (15). It is thus surprising to find a positive relationship between HB BOLD activity and reward magnitude (a). We therefore performed further analyses to investigate the relationship between HB activity, reward prospect, and action time. Interestingly, we found that the positive relationship between HB BOLD signal and action time (Fig.3) is mainly driven by trials on which participants were expecting to lose (on trials where rate of change in reward probability was negative) (g). This finding corresponds with the previous literature that HB is more active when a loss or an aversive event is predicted. The lines and shadings show the mean and standard error of the β weights across the participants, respectively. Time zero is the response time. Significance testing on time-course data was performed by using a leave-one-out procedure on the group peak signal. One-sample t-tests with Holm-Bonferroni correction. * P<0.05, ** P<0.01, *** P<0.001.  Fig.3B). However, we found no significant relationship between EV and BOLD activity (b,d). Format as in Figure 3.    basalis. However, the BF mask that we used in this study mainly covers medial septal and diagonal nuclei. The hippocampus (HPC), however, provides a major input to the septal nuclei and septal nuclei projects back to the hippocampus. We therefore updated our model ( Fig.7) by adding the BOLD activity from hippocampus to the pathway between ACC and BF (a). We found a significant path from ACC to hippocampus (b=-0.17, P<0.001) and from hippocampus to BF (b=0.21, P<0.001). We then compared the updated model with an alternative model with identical number of degrees of freedom but with the direction of paths reversed (b). The updated model (AIC=474288) performed better than the control model (AIC=583003). However, the updated model was no better than our original model (AIC=377600) at explaining the data.