Data-driven Selection of Relevant fMRI Voxels using Sparse CCA and Stability Selection

Functional MRI data is both high-dimensional and noisy, and consequently many machine learning methods overfit severely if applied without first reducing the dimensionality. We propose an algorithm to find the most relevant voxels automatically using a Sparse Canonical Correlation Analysis (SCCA) model, the statistical framework of stability selection, and a rich feature representation for the experimental task. We empirically validate the voxels chosen by our algorithm by applying it to fMRI data from 16 subjects listening to classical music classified as’ happy’,’sad’or’neutral’. Our algorithm selects voxels based on automatically extracted music features, without access to the label information. We show that a linear Support Vector Machine classifier trained on a small subset of voxels selected using our method can match or beat an SVM trained on the full brain in the task of predicting the labels’ happy’and’sad’, while performing much better than an SVM trained on a similar randomized voxel set. We further show using Representational Similarity Analysis that the voxel subsets chosen by our method are related to voxels that are denoted as active by traditional Statistical Parametric Mapping.