DIAGNOSTIC CLASSIFICATION OF ASD IMPROVES WITH DYNAMIC FC OF FMRI COMPARED TO STATIC FC

Abstract

In our study, we compared the static functional connectivity (sFC) and dynamic functional connectivity (dFC) computed from functional magnetic resonance imaging (fMRI) to diagnose individuals with autism spectrum disorder (ASD) and typically developing (TD). We collected fMRI data of 112 samples (57 TD and 55 ASD) from the ABIDE databases, preprocessed, and extracted time series data based on blood-oxygen-level-dependent signals from 236 regions of interest across cortical, subcortical, and cerebellar areas, as defined by Gordon's, Harvard Oxford, and Diedrichsen atlases. We divided the time series data of each sample into six windows, each with a 30-second, resulting in a total of 672 samples. Further, Pearson correlations were computed between the regions from 180 seconds (sFC) and each 30-second window time series (dFC), generating 236x236 matrices and considering the upper/lower triangular matrix, leading to 27,730 features for each sample. Subsequently, we ranked the features using the extreme Gradient Boost (XGBoost) method and fed the selected features to various machine learning classifiers, including logistic regression (LR), support vector machine, multi-layer perceptron (MLP), and XGBoost. The machine learning models' performance was evaluated using five-fold crossvalidation, and accuracy, sensitivity, specific-ity, precision, and f1-score were assessed. Our results revealed that sFC produced a classification accuracy of 88.76% using the MLP classifier. The classification accuracy improved to 96.65% using the dFC and LR classifier with the top 1900 features. Our results show that dFC is able to improve the classification accuracy of ASD diagnosis compared to sFC. © 2024 IAE All rights reserved.