Machine-learning classification of children and adolescents with ASD using resting-state frequency-specific intrinsic activity.
Authors
Affiliations (4)
Affiliations (4)
- MOE Key Laboratory for Neuroinformation, Center for Information in Medicine, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China.
- High-Field Magnetic Resonance Brain Imaging Key Laboratory of Sichuan Province, Center for Information in Medicine, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China.
- China-Cuba Belt and Road Joint Laboratory on Neurotechnology and Brain-Apparatus Communication, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China.
- Department of Biomedical Engineering, New Jersey Institute of Technology, Newark, NJ, United States.
Abstract
Autism spectrum disorder (ASD) is characterized by heterogeneous developmental trajectories, yet it remains unclear whether frequency-specific resting-state functional magnetic resonance imaging (rs-fMRI) features can distinguish age-defined developmental stages within the condition. We analyzed rs-fMRI data from 251 participants with ASD, comprising 146 children and 105 adolescents aggregated from ten sites in the Autism Brain Imaging Data Exchange (ABIDE). ALFF and ReHo were computed across three frequency bands: Conventional (0.01-0.08 Hz), slow-4 (0.027-0.073 Hz), and slow-5 (0.01-0.027 Hz). Region-of-interest features were extracted using the 246-region Brainnetome Atlas. To ensure rigorous generalization, participants were divided into a stratified training set (80%, <i>n</i> = 200) and a held-out test set (20%, <i>n</i> = 51), with stratification based on the child-adolescent group label and a fixed random seed of 42. CovBat harmonization parameters, feature-scaling parameters, LASSO feature selection, and classifier hyperparameters were estimated using the training data only and subsequently applied to the held-out test data. Final model performance was evaluated once on the held-out test set. Performance was evaluated using Logistic Regression (LR), Support Vector Machine, and Random Forest classifiers, with Shapley Additive Explanations (SHAP) used to characterized interpret feature contributions. The slow-4 and Conventional-band features showed higher held-out ASD test-set performance than slow-5 features. The best single-metric model by area under the receiver operating characteristic curve (AUC) was slow-4 ReHo Logistic Regression, which achieved an AUC of 0.811 and accuracy of 0.745. The exploratory combined model using slow-4 ALFF and ReHo features achieved the highest overall AUC of 0.819 (accuracy = 0.725). SHAP analysis identified distributed model-contributing regions in the slow-4 ReHo model, including the inferior parietal lobule, lateral occipital cortex, middle and inferior frontal gyri, basal ganglia, and thalamus. Frequency-specific resting-state features, particularly local synchronization in the slow-4 band, capture developmental-stage-related variation within ASD. The involvement of frontoparietal, visual, and subcortical networks suggests that developmental heterogeneity in ASD is supported by distributed reorganization of intrinsic brain activity. These findings highlight the potential of frequency-specific rs-fMRI metrics as candidate markers for characterizing neurodevelopmental stages in ASD, warranting further validation in longitudinal and independent cohorts.