The Classification System and Biomarkers for Autism Spectrum Disorder - A Machine Learning Approach.

Autism spectrum disorder (ASD) is a kind of neurodevelopmental disorder. ASD patients are usually difficult in social communication and daily life, and there are no drugs to cure ASD. A large number of clinical cases exhibits an early intervention is beneficial for ASD patients. Therefore, the rapid and accurate diagnosis is a great of significance. With the development of artificial intelligence techniques, machine learning (ML) models have been applied in analyzing the ASD. However, it is an extremely challenging task to reveal the biomarker through ML models. In this paper, a computational protocol is proposed to differentiate ASD from typical development (TD) using the resting-state functional magnetic resonance imaging (rs-fMRI) images of brains and reveal the brain regions related to ASD. The computational protocol is consisted of feature selection, model training, and feature analysis. Classification models are constructed based on XGBoost algorithm, and show a better performance compared with previous well-known models. By analyzing the input features of models, the functional connections of cingulo-opercular network (CON) and default-mode network (DMN) is founded to contribute the models' performance significantly, which can be regarded as the biomarker of ASD.