Can neuroimaging-based machine learning deliver reliable ADHD classification? A systematic review of classification studies

Attention-deficit/hyperactivity disorder (ADHD) diagnosis currently relies primarily on subjective clinical reports, prompting investigation into more objective, brain-based markers using neuroimaging and machine learning. This systematic review examined 236 studies employing machine learning techniques on electroencephalography (EEG) and magnetic resonance imaging (MRI) data for ADHD classification. Overall reported classification accuracy across studies was high (87%), with a substantial upward trend over publication years. We found higher classification accuracy in studies using EEG, those that included a cognitive task during data acquisition, and those focusing on younger participants. However, most studies compared ADHD with typically developing participants, and only a very small number included other clinical comparison groups, which may be more informative for clinical translation. In addition, smaller studies tended to report higher accuracies than larger studies, and many studies used validation procedures that may be more prone to overfitting, underscoring the need for larger, more controlled studies in future work. These findings suggest that neuroimaging based machine learning might support reliable ADHD classification under specific methodological conditions, while also highlighting substantial methodological limitations and heterogeneity in the literature that currently constrain direct clinical translation.