Articles
Network Signatures of Verbal Ability in Autism: A Multisite fMRI Study Using Graph Theory and Machine Learning
Download
Authors
-
Huibin Ma
School of Information and Electronics Technology, Jiamusi University, Jiamusi 154007, China
-
Jinying Wu
School of Information and Electronics Technology, Jiamusi University, Jiamusi 154007, China
-
Lanfen Chen
School of Medical Imaging, Shandong Second Medical University, Weifang 261053, China
-
Jiaying Song
School of Information and Electronics Technology, Jiamusi University, Jiamusi 154007, China
-
Hang Zhang
School of Medical Imaging, Shandong Second Medical University, Weifang 261053, China
-
Zeqi Hao
School of Psychology, Zhejiang Normal University, Jinhua 321004, China
-
Linlin Zhan
Faculty of Western Languages, Heilongjiang University, Harbin 150080, China
-
Lulu Cheng
School of Foreign Studies, China University of Petroleum (East China), Qingdao 266580, China
-
Xize Jia
School of Psychology, Zhejiang Normal University, Jinhua 321004, China
Intelligent Laboratory of Zhejiang Province in Mental Health and Crisis Intervention for Children and Adolescents, Zhejiang Normal University, Jinhua 321004, China
Received: 3 August 2025; Revised: 15 October 2025; Accepted: 17 December 2025; Published: 2 January 2026
Autism spectrum disorder (ASD) is associated with atypical functional connectivity (FC), but its relationship with verbal ability remains unclear. In this study, resting-state functional magnetic resonance imaging (rs-fMRI) data from 219 individuals with ASD and 322 healthy controls (HCs) with valid verbal intelligence quotient (VIQ) scores were obtained from 12 sites in the Autism Brain Imaging Data Exchange (ABIDE) database. Region of interest (ROI)-based FC and whole-brain network metrics were computed and correlated with VIQ. Three feature sets—raw connectivity and topology, group-difference metrics, and group-difference metrics combined with VIQ—were evaluated using support vector machine (SVM) and manifold learning. ASD showed widespread FC alterations with overall increases and localized decreases in specific networks. At the global level, ASD patients exhibited a significantly lower area under the receiver operating characteristic curve (AUC) specifically for the normalized clustering coefficient γ compared with HCs. Nodal analyses revealed increased degree centrality and efficiency in subcortical and frontal regions, but decreased values in limbic and temporal regions in ASD. Centrality of the anterior cingulate cortex showed a negative correlation with VIQ, while nodal efficiency and degree centrality in occipital regions showed positive correlations. Classification using the raw feature set achieved the best overall performance under the Principal Component Analysis-Uniform Manifold Approximation and Projection-Support Vector Machine (PCA-UMAP-SVM) pipeline, whereas group-difference features augmented with VIQ yielded the highest AUC among the standard SVM models. These results suggest that VIQ-anchored network features may help explain variability in verbal ability and brain organization in ASD.