Classification of Autism Spectrum Disorder in Children Using Electroencephalography Power Ratios Obtained During a...

Autism spectrum disorder (ASD) diagnosis relies on behavioral observation, but a shortage of qualified experts leads to delayed diagnosis, with the average age of diagnosis being 4.8 years. In this study, we aimed to assess the ability of resting-state electroencephalography (EEG) power spectral features and EEG features during naturalistic theory of mind (ToM) tasks to distinguish children with ASD from typically developing (TD) children and to evaluate early screening potential.

A cross-sectional diagnostic study was conducted among 183 Chinese children ages 3 to 11 years (83 with physician-diagnosed ASD and 100 TD children). After quality control, the EEG data of 163 participants were analyzed. Participants wore EEG devices while they watched Disney's "Partly Cloudy" as a naturalistic social task and completed the resting-state recordings. The primary outcome was XGBoost-based ASD classification performance using resting-state EEG power spectral features and EEG features, evaluated by accuracy, area under the curve (AUC), sensitivity, and precision.

A total of 163 participants (73 ASD, 90 TD) were analyzed. The groups differed significantly in sex (male proportion: 89.15% vs. 67.00%, p < .001) and IQ (92.85 vs. 112.43, p = .035). The mental-control power ratio model performed best, with an accuracy of 0.925 (95% CI, 0.909-0.940) and an AUC value of 0.980 (95% CI, 0.972-0.986). The performance of the resting-state models was poor (accuracies: 0.549 and 0.515). Cross-age prediction remained robust, with accuracies of ∼90% to 92% and AUCs >97%, showing only slightly reduced precision in the youngest group.

Unlike resting-state EEG features, EEG power ratios during naturalistic ToM tasks distinguish children with ASD from TD children with high accuracy.