Machine learning methods for autism spectrum disorder classification

Abstract The use of AI-driven predictive models to identify patterns that can act as biomarkers for different neuropathological conditions is becoming highly prevalent. In particular, the recent advances in representation learning techniques provide unprecedented opportunities to understand, diagnose, and eventually treat a gamut of neurological conditions, including epilepsy, stroke, and autism. In this chapter, we consider the challenging problem of autism spectrum disorder classification using resting state fMRI (rs-fMRI) measurements. While supervised learning methods provide a convenient way to build predictive functions that can identify specific signatures in rs-fMRI pertinent to different levels of severity, the key challenge is in modeling the complex spatiotemporal data (from brain networks) and leveraging the dependencies among patients in a population. In this context, graphs provide a powerful framework to address both these challenges and enable the design of sophisticated predictive modeling tools. The advances in graph signal analysis and the generalization of machine learning techniques, such as kernel methods and deep neural networks, to graphs, make them suitable for challenging diagnosis problems. In this chapter, we describe how graphs, either at patient level or population level, can be used to build predictive models. We present mathematical formulations, algorithmic solutions, and empirical studies to establish graph-based machine learning as an effective solution for autism diagnosis.