Autism spectrum disorder diagnosis using fractal and non-fractal-based functional connectivity analysis and machine learning methods

Autism spectrum disorder (ASD) is a neurological condition characterized by impaired functional connectivity (FC) networks in the brain. There are several brain networks associated with ASD that have been studied for ASD diagnosis, but the results are inconsistent. A functional magnetic resonance imaging (fMRI) study was performed to address this gap by comparing brain networks among autistic individuals and individuals with typical development (TD) using data from the ABIDE-I and ABIDE-II databases. Blood oxygen level-dependent (BOLD) time series were extracted from 236 regions of interest (ROI) in fMRI data using three atlases: Gordon’s, Harvard Oxford, and Diedrichsen. Consequently, 27,730 nonlinear features are extracted from FC matrices, including fractals, non-fractals, and Pearson correlation coefficients (PCC). A parametric and nonparametric classifier was used to analyze the top 0.1%, 0.3%, 0.5%, 0.7%, 1%, 2%, and 3% of features based on the XGBoost feature ranking algorithm. In the study, we found that non-fractal brain FC measures can accurately identify ASD and TD more effectively than fractal and PCC measures. Classifiers performed well, with FC features at the top 0.3%. The classification model at OHSU was more accurate than the model at other sites. There was 100% accuracy at a single site and 96.17% accuracy at all sites using a multilayer perceptron classifier with non-fractal features. The classifier model shows that Cingulo-Parietal Task Control (13.6%), Retro-Splenial Temporal (RST) (13.3%), and Salience (11.6%) are significant contributors. The optimal performance was observed in features derived from networks such as RST and default (4 connections), auditory and Fronto-Parietal Task Control (3 connections), Cingulo-Opercular Task Control (COTC), and ventral attention (3 connections) within COTC (3 connections), visual and COTC (3 connections), and cerebellum and COTC (3 connections). Based on the results, non-fractal-based FC has value in distinguishing ASD from TD using resting-state fMRI.