Abstract

Discriminative supervised learning algorithms, such as Support Vector Machines, are becoming increasingly popular in biomedical image computing. One of their main uses is to construct image-based prediction models, e.g., for computer aided diagnosis or "mind reading." A major challenge in these applications is the biological interpretation of the machine learning models, which can be arbitrarily complex functions of the input features (e.g., as induced by kernel-based methods). Recent work has proposed several strategies for deriving maps that highlight regions relevant for accurate prediction. Yet most of these methods o n strong assumptions about t he prediction model (e.g., linearity, sparsity) and/or data (e.g., Gaussianity), or fail to exploit the covariance structure in the data. In this work, we propose a computationally efficient and universal framework for quantifying associations captured by black box machine learning models. Furthermore, our theoretical perspective reveals that examining associations with predictions, in the absence of ground truth labels, can be very informative. We apply the proposed method to machine learning models trained to predict cognitive impairment from structural neuroimaging data. We demonstrate that our approach yields biologically meaningful maps of association.