A Systematic Assessment of Feature Extraction Methods for Robust Prediction of Neuropsychological Scores from Functional Connectivity Data

Multivariate prediction of human behavior from resting state data is gaining increasing popularity in the neuroimaging community, with far-reaching translational implications in neurology and psychiatry. However, the high dimensionality of neuroimaging data increases the risk of overfitting, calling for the use of dimensionality reduction methods to build robust predictive models. In this work, we assess the ability of four dimensionality reduction techniques to extract relevant features from resting state functional connectivity matrices of stroke patients, which are then used to build a predictive model of the associated language deficits based on cross-validated regularized regression. Features extracted by Principal Component Analysis (PCA) were found to be the best predictors, followed by Independent Component Analysis (ICA), Dictionary Learning (DL) and Non-Negative Matrix Factorization. However, ICA and DL led to more parsimonious models. Overall, our findings suggest that the choice of the dimensionality reduction technique should not only be based on prediction/regression accuracy, but also on considerations about model complexity and interpretability.