Project 464461

Autism spectrum disorder (ASD) refers to a set of neurodevelopmental disorders marked by stereotyped behaviours and impairments in social interaction with lifelong impacts. Its complex etiology and heterogeneous presentation renders diagnosis a substantial clinical challenge that currently relies on longitudinally gathering often variably reliable and inaccessible behavioral measures, even though earlier diagnosis is associated with better treatment outcomes. Thus, there is a pressing need for automated reliable diagnostic and monitoring tools for ASD that are patient-specific. Currently, most automated ASD detection methods rely on computationally-efficient machine-learning (ML) models like deep artificial neural networks (ANNs) for establishing links between specific patterns of whole-brain resting-state fMRI features and clinical variables. However, the high-dimensionality of fMRI data impedes the accuracy and thus clinical usefulness of ML methods. At the same time, advances in computational neuroscience are enabling the construction of large-scale brain networks, which can realistically simulate relevant neural processes implicated in ASD pathophysiology; but these models lack methods to efficiently be tailored to single patients' measured fMRI data to be incorporated into diagnosis or monitoring. Based on our previous research, we hypothesise that the integration of machine learning and biophysical modeling can create a clinically usable classification tool for ASD.