Data-driven learning strategies for biomarker detection and outcome prediction in Autism from task-based fMRI

Autism spectrum disorder (ASD) is a human developmental disorder that affects how people interact and behave, with impaired social interaction and communication, as well as repetitive behaviors, with severity ranging from mild to significantly disabling. The prevalence continues to rise and the associated costs are enormous. In this chapter, we present advanced methods for the analysis of task-based functional magnetic resonance imaging (fMRI) for classification and characterization of individuals, facilitating the identification of ASD imaging biomarkers and personalized outcome prediction for behavioral therapy used for ASD. Included are descriptions of our deep learning techniques for extracting biomarkers and predicting outcome via novel strategies focusing on spatial characteristics using Graph Neural Networks (GNNs) as well as temporal characteristics with Long Short Term Memory (LSTM) networks. Improved use and characterization of the dynamic changes in connectivity appear crucial for advancing performance based both on our work and the literature. Thus, in addition, we describe our efforts aimed toward the development of a richer, integrated model that can more fully exploit the complete spatio-temporal characteristics of the data and its inherent dynamics based on temporally-windowed strategies as well as through the use of an advanced model of causality based on the solution of neural ordinary differential equations (ODEs).