Bayesian deconvolution for computational cognitive modeling of fMRI data

A central goal of cognitive neuroscience is to make inferences about underlying cognitive processes from observable data. However, current fMRI analysis tools cannot directly estimate latent parameters in computational cognitive models from blood-oxygen-level-dependent (BOLD) signal. Here, we present a novel Bayesian deconvolution technique for full hierarchical generative cognitive modeling of fMRI timeseries data. We validated this approach by applying Bayesian deconvolution to the monetary incentive delay (MID) task to identify processes underlying incentive anticipation in a sample of 54 individuals who underwent 2 scan sessions as part of a clinical trial for anxiety and depression. Based on a series of Bayesian models, we found evidence that striatal reward region activity reflects incentive prediction error rather than raw incentive value during anticipation of monetary loss or gain. Test-retest analyses found that individual parameters estimated using a generative Bayesian learning model (including a persistent prior parameter and a β parameter representing a scaling term between prediction error and BOLD signal) were estimated more reliably than an index derived from traditional fMRI analysis (beta value for contrast between gain and no gain during anticipation). Our method holds potential for broad application to diverse neural processes and individual differences in health and disease.