Indirect decoding of behavior from brain signals via reconstruction of experimental conditions

Abstract

The data acquired during a functional magnetic resonance imaging (fMRI) experiment usually comprise experimental conditions, brain signals and behavioral responses. This reflects the underlying causal flow where the experimental conditions evoke brain responses that in turn result in behavior. In multivariate analyses, a common approach is to focus on the second step of that chain and to decode behavioral responses from brain signals. However, a different approach would be to first reconstruct the experimental conditions from brain signals and in turn use these to predict behavior. While this indirect approach would go against the causal chain of events, it might work better under certain circumstances, especially when the experimental conditions evoke much stronger measurable brain signals than the overt motor behavior. Here we tested this question directly by assessing the various mappings between conditions, brain signals and behavior in an open dataset. We found that the path of first decoding experimental conditions works surprisingly well, even though in our example data set, it is still outperformed by directly decoding the behavior from brain responses.