Autism Spectrum Disorder Diagnosis Using Sparse Graph Embedding of Morphological Brain Networks

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder involving a complex cognitive impairment that can be difficult to diagnose early enough. Much work has therefore been done investigating the use of machine-learning techniques on functional and structural connectivity networks for ASD diagnosis. However, networks based on the morphology of the brain have yet to be similarly investigated, despite research findings that morphological features, such as cortical thickness, are affected by ASD. In this paper, we first propose modelling morphological brain connectivity (or graph) using a set of cortical attributes, each encoding a unique aspect of cortical morphology. However, it can be difficult to capture for each subject the complex pattern of relationships between morphological brain graphs, where each may be affected simultaneously or independently by ASD. In order to solve this problem, we therefore also propose the use of high-order networks which can better capture these relationships. Further, since ASD and normal control (NC) high-dimensional connectomic data might lie in different manifolds, we aim to find a low-dimensional representation of the data which captures the intrinsic dimensions of the underlying connectomic manifolds, thereby allowing better learning by linear classifiers. Hence, we propose the use of sparse graph embedding (SGE) method, which allows us to distinguish between data points drawn from different manifolds, even when they are too close to one another. SGE learns a similarity matrix of the connectomic data graph, which then is used to embed the high-dimensional connectomic features into a low-dimensional space that preserves the locality of the original data. Our ASD/NC classification results outperformed several state-of-the-art methods including statistical feature selection, and local linear embedding methods