<![CDATA[Autism spectrum disorder (ASD) is primarily identified through behavioral and developmental assessment, which may require specialist expertise and considerable time. Electroencephalography (EEG) offers non-invasive neurophysiological information, but reliable ASD classification remains challenging because EEG signals are nonstationary and highly variable across individuals. This study evaluates whether visibility-graph topology derived from EEG power spectral density (PSD) curves provides complementary information beyond conventional spectral descriptors. Resting-state EEG from the Banda Aceh dataset (7 ASD and 7 typically developing participants, 250 Hz) and the KAU dataset (8 ASD and 8 control participants, 256 Hz) were processed independently. Preprocessing included quality inspection, 0.5–45-Hz zero-phase filtering, dataset-specific 50/60-Hz notch filtering, common average referencing, and non-overlapping 30-s segmentation. Welch PSD was estimated using 2-s Hann windows with 50% overlap across five frequency bands. Direct PSD, natural visibility graph features, and horizontal visibility graph features were compared using RBF-SVM and a six-hidden-layer MLP. Evaluation used subject-wise balanced cross-validation, with performance calculated at the epoch level from participant-disjoint outer-test folds. VG-NVG-PSD with SVM achieved the best observed accuracy of 97.14 ± 3.29% on Banda Aceh, representing a 0.97-percentage-point numerical gain over direct PSD-SVM. VG-HVG-PSD with SVM achieved 96.83 ± 2.34% on KAU, representing a 0.35-percentage-point gain. SVM produced the highest observed configuration on both datasets, whereas graph-spectral transformation did not consistently improve MLP performance. Bonferroni-corrected pairwise comparisons did not establish statistical superiority among individual methods. These findings indicate that PSD-derived visibility graphs provide an interpretable complementary representation for EEG-based ASD classification, although the preferred visibility criterion was dataset-dependent. Larger cohorts and external validation are required before clinical generalization]]>
