<![CDATA[This study addresses a key clinical challenge in EEG-based autism spectrum disorder (ASD) diagnosis: most prior work reports accuracy at the window level, whereas the clinical diagnostic unit is the individual subject, a mismatch that can inflate performance estimates. We develop and validate a subject-level EEG feature aggregation framework, combining Wavelet Packet Transform (WPT) energy-variance features, fold-safe ANOVA feature selection, and per-subject aggregation, evaluated under leakage-safe nested leave-one-subject-out (LOSO) cross-validation. Sixteen-channel EEG was recorded from 16 participants (8 ASD, 8 neurotypical) in Banda Aceh, Indonesia, using a single OpenBCI Cyton board with a Daisy module for native simultaneous 16-channel acquisition, preprocessed with Infomax ICA artifact removal, and classified using Logistic Regression. The framework achieved a nested LOSO balanced accuracy of 93.75% (F1 = 0.9412); because this estimate derives from only 16 subjects, the repeated subject-pair evaluation (balanced accuracy 0.9525 ± 0.0529, AUC 0.9897 ± 0.0163, permutation p = 0.0196) is treated as the primary robustness indicator. A learning-curve analysis indicated that performance approaches, but has not fully reached, a plateau at this cohort size. Sensitivity analyses across nine window/overlap configurations and six wavelets, selected within the nested cross-validation, confirmed the result was not selection-dependent. A controlled ablation indicated Infomax ICA outperformed MSICA-DWT (Wilcoxon p < 0.001), though a participant-level bootstrap interval was wide and included zero, reflecting the limited power of the 16-subject cohort. Overall, these results demonstrate the methodological feasibility of subject-level WPT energy-variance aggregation for leakage-safe ASD screening research rather than diagnostic readiness, and external multicenter validation on larger, demographically characterized cohorts is required before any clinical use.]]>
