Characterizing and Identifying Autism Disorder Using Regional Connectivity Patterns and Extreme Gradient Boosting Classifier

Abstract

In this paper, we design and implement a procedure to capture and extract regional connectivity patterns from brain connectomics. Moreover, we assess the viability of such patterns as predictors for both childhood and adult autism. Finally, we investigate which regions and connections are significant for characterizing and predicting this psychiatric pathology. We use two publicly-available neuroimaging datasets and systematically train 90 extreme gradient boosting trees classifiers (XGBoost) for each set, each classifier receiving connectivity patterns extracted for one of the 90 regions of interest that form the automated anatomical labeling (AAL) atlas. Our most predictive regional connectivity pattern features achieved an accuracy of 78.95% (precision = 78.98%, recall = 78.75%) for the adult population and 75.01% accuracy for the pediatric dataset (precision = 75.00%, recall = 75.09 %) for the pediatric population. These classification accuracies are higher than those reported in prior studies that used the same datasets. Altogether, our results indicate that local connectivity around the lingual gyrus can predict both adult and childhood autism with relatively high accuracy.

This study was supported by the National Natural Science Foundation of China (Projects No. 61572239 and No. 61772242), and the Doctoral Fund of the Ministry of Education of China (Project No. 2017M611737).