Within-cohort discrimination and cross-cohort transfer of EEG-based ADHD classifiers: An exploratory two-cohort evaluation

We evaluated whether participant-separated within-cohort discrimination was retained when electroencephalography (EEG) classifiers for attention-deficit/hyperactivity disorder (ADHD) were applied to a different population and recording task. This exploratory reanalysis included 121 children performing a visual task and 79 adults recorded during eyes-open rest after basic channel screening. Both cohorts used Cz/F4, 128 Hz, and the first fifteen non-overlapping 2-second epochs per participant. Fixed logistic regression and random forest classifiers used spectral, amplitude-derived, and combined features. Internal evaluation used five repetitions of five-fold participant-grouped cross-validation; bidirectional transfer fitted transformations and classifiers only on the source cohort. Adult spectral balanced accuracy (BA) was 0.9519-0.9546, compared with 0.5504-0.5664 in children. Adult spectral logistic regression exceeded all 999 internal label permutations (Monte Carlo p=0.001; within-family Holm-adjusted p=0.002). Paired spectral-minus-amplitude conditional intervals excluded zero in adults but spanned zero in children. Cross-cohort spectral BA was 0.3584-0.4675, with area under the receiver operating characteristic curve (AUC) of 0.0309-0.3857 in the unchanged score direction. Increasing pediatric source-epoch budgets moved random-forest AUC toward 0.5, whereas logistic-regression AUC remained below 0.1. Strong adult internal discrimination thus coexisted with reverse cross-cohort ranking. These findings characterize the evaluated pipelines, not clinical diagnostic utility or a unique mechanism: age, task, and acquisition differences were not separated, target results had previously been inspected, and conditional intervals excluded model-refitting uncertainty.