Abstract
In contrast to the method of constant stimuli, adaptive procedures dynamically select the next stimulus based upon the prior responses of the subject, and an underlying stimulus-response model. For fMRI experiments that seek to measure the parameters of a neural response function, Bayesian adaptive stimulus selection (such as provided by QUEST+; Watson 2017) may provide better estimates of model parameters than traditional approaches. QUEST+ may be particularly useful in cases where the neural response is a function with multiple parameters, which must be fit simultaneously. Despite these advantages to adaptive stimulus selection, a complication is that QUEST+ considers responses as the proportion of outcomes within pre-defined, discrete bins, while BOLD fMRI data is a continuous signal with an uncertain amplitude and a varying baseline. To account for this, we have implemented a framework for QUEST+ fMRI that includes: 1) initial model fitting to the time-series to extract a gain parameter for each stimulus event, accounting for the hemodynamic response; 2) a procedure for dynamically updating the mapping between responses relative to a reference stimulus, and the fixed set of outcomes specified by QUEST+; 3) a Gaussian noise parameter that intercedes between the parameterized model of response and the proportions of observed outcomes. We examined the measurement of the V1 cortical response to a high-contrast stimulus flickering at different frequencies, as fit by a 4-parameter difference-of-exponential temporal sensitivity model. In simulations that model empirical data, we find that the QUEST+ approach recovers model parameters more accurately than does constant stimuli given a fixed acquisition length. This framework now allows us to optimize experimental design (e.g., the number and duration of stimulus types) given a neural response function to be measured.