A deep learning approach for feature extraction from resting state functional connectivity of stroke patients and prediction of neuropsychological scores

Deep learning models are being increasingly used in precision medicine thanks to their ability to provide accurate predictions of clinical outcome from large-scale datasets of patient’s records. However, in many cases data scarcity has forced the adoption of simpler (linear) feature extraction methods, which are less prone to overfitting. In this work, we exploit data augmentation and transfer learning techniques to show that deep, non-linear autoencoders can in fact extract relevant features from resting state functional connectivity matrices of stroke patients, even when the available data is modest. In particular, we used the Human Connectome Project (HCP) which is a large and high-quality dataset to learn latent representation of healthy patients. The latent representations extracted by the autoencoders can then be given as input to regularized regression methods to predict neurophsychological scores, outperforming recently proposed methods based on linear feature extraction. Additionally, we study the impact of the cross validation set-up for each model, and we examined the quality of the predictive maps obtained by back-projecting the regression weight, to display the most predictive RSFC edges.