Task-driven ICA feature generation for accurate and interpretable prediction using fMRI.

Functional Magnetic Resonance Imaging (fMRI) shows significant potential as a tool for predicting clinically important information such as future disease progression or drug effect from brain activity. Multivariate techniques have been developed that combine fMRI signals from across the brain to produce more robust predictive capabilities than can be obtained from single regions. However, the high dimensionality of fMRI data makes overfitting a significant problem. Reliable methods are needed for transforming fMRI data to a set of signals reflecting the underlying spatially extended patterns of neural dynamics. This paper demonstrates a task-specific Independent Component Analysis (ICA) procedure which identifies signals associated with coherent functional brain networks, and shows that these signals can be used for accurate and interpretable prediction. The task-specific ICA parcellations outperformed other feature generation methods in two separate datasets including parcellations based on resting-state data and anatomy. The pattern of response of the task-specific ICA parcellations to particular feature selection strategies indicates that they identify important functional networks associated with the discriminative task. We show ICA parcellations to be robust and informative with respect to non-neural artefacts affecting the fMRI series. Together, these results suggest that task-specific ICA parcellation is a powerful technique for producing predictive and informative signals from fMRI time series. The results presented in this paper also contribute evidence for the general functional validity of the parcellations produced by ICA approaches.