A comparison of fMRI and behavioral models for predicting inter-temporal choices

In an inter-temporal choice (IteCh) task, subjects are offered a smaller amount of money immediately or a larger amount at a later time point. Here, we are using trial-by-trial fMRI data from 363 recording sessions and machine learning in an attempt to build a classifier that would ideally outperform established behavioral model given that it has access to brain activity specific to a single trial. Such methods could allow for future investigations of gatelike factors that influence IteCh choices. To investigate this, coefficients of a GLM with one regressor per trial were used as features for a support vector machine (SVM) in combination with a searchlight approach for feature selection and cross-validation. We then compare the results to the performance of four different behavioral models. We found that the behavioral models reached mean accuracies of 90% and above, while the fMRI model only reached 54.84% at the best location in the brain with a spatial distribution similar to the well-known value-tracking network. This low, though significant, accuracy is in line with simulations showing that classifying based on signals with realistic correlations with subjective value produces comparable, low accuracies. These results emphasize the limitations of fMRI recordings from single events to predict human choices, especially when compared to conventional behavioral models. Better performance may be obtained with paradigms that allow the construction of miniblocks to improve the available signal-to-noise ratio.