Temporal and Cross-Subject Probabilistic Models for fMRI Prediction Tasks

We present a probabilistic model applied to the fMRI video rating prediction task of the Pittsburgh Brain Activity Interpretation Competition. Using a Dynamic Gaussian Markov Random Field, we model the relationship between the subjects’ fMRI voxel measurements and the rated properties of the videos, such as presence of language or subject amusement. Also included in our model are dependencies of the ratings across time steps, and between subjects. Rather than grouping voxels into regions, we chose to use individual voxels as features. However, for some of the ratings, we found that using a prior with a bias toward similar parameters for neighboring voxels improved our predictions. Our model performed significantly better than a baseline regression model on held out training data, and displayed good performance in predicting the scored ratings across the three subjects in the training and test data sets. We chose a general model that is applicable to all the rating categories, rather than specially attacking single rating types at a time.