Network-Guided Group Feature Selection for Classification of Autism Spectrum Disorder

We present an anatomically guided feature selection scheme for prediction of neurological disorders based on brain connectivity networks. Using anatomical information not only gives rise to an interpretable model, but also prevents overfitting, caused by high dimensionality, noise and correlated features. Our method selects meaningful and discriminative groups of connections between anatomical regions, which can be used as input for any supervised classifier, such as logistic regression or a support vector machine. We demonstrate the effectiveness of our method on a dataset of autism spectrum disorder, with an AUC of 0.76, outperforming baseline methods.