Multivariate Prediction of Human Behavior in Task fMRI

Alessandra Camassa, Joseph Park, Margot Wagner, Terrence Sejnowski, Gerald M Pao

Keywords: human behavior, time series prediction, dynamical systems, Takens theorem, dimensionality reduction, explainability, task fMRI

TL;DR: A new algorithm for the prediction of human behavior in task fMRI data using state space reconstruction and causal inference.

Abstract: A central problem of research at the intersection between neuroscience, machine learning and artificial intelligence is to understand and model the relationship between brain activity and behavior in humans. Here we propose an algorithm able to identify the brain regions that are causally relevant to a specific behavior and use them to build a predictive model for time series data. Our goal is to predict the behavioral output of human subjects executing a task within the fMRI scan using the hemodynamic response of different brain regions as an indirect measure of neural activity. We use the Simplex prediction algorithm and a new iterative feature evaluation method: this approach is based on state space reconstruction, and the variables chosen to build a low dimensional representation of the behavior are selected according to their relationship to the target time series (i.e. the behavioral time series). To discover these relationships, we use convergent cross mapping (CCM), a Takens theorem based causal inference method that can detect nonlinear causation between time series. We iteratively select the variables that enhance forecasting skill and with convergent causal coupling to the target to construct a multivariate local manifold able to predict the behavior through multivariate Simplex projection. This method facilitates the integration of empirical observations of brain activity into forecasting of realistic behaviors. Our approach allows to keep a direct correspondence between observations and model variables while optimizing the prediction accuracy.

Submission Number: 19