Attention-deficit/hyperactivity disorder classification in functional magnetic resonance images with deep learning

Attention-deficit/hyperactivity disorder (ADHD) affects about 5–7% of children. Its subtypes, characterized by distinct patterns of brain alterations, require distinct medications and treatments. This study aims to develop an ADHD classifier based on functional magnetic resonance imaging (fMRI) that is capable of distinguishing between combined (ADHD-C) and inattentive (ADHD-I) subtypes while also identifying relevant neural biomarkers associated with the disorder. This research employed ensemble learning strategies using convolutional neural networks (CNNs) trained on resting-state fMRI data of 623 participants. Ensemble methods use CNN predictions based on connectivity networks as features to generate a final classification. This study used the following ensemble methods: majority voting, linear ridge classifier, adaptive boosting (AdaBoost) and extreme gradient boosting (XGBoost). The majority voting and XGBoost classifiers achieved accuracies of 97.6 ± 2.1% for ADHD diagnosis and 94.8 ± 1.3% for subtype classification in the cross-validation, respectively. The limbic and somatomotor connectivity networks were identified as reliable biomarkers for determining the ADHD subtype. For the diagnosis of ADHD, the visual, somatomotor, frontoparietal, and limbic networks have shown prominence as biomarkers. As far as we know, this is the largest study to date on ADHD among those that have achieved over 77% accuracy in subtype classification. The findings of this study advance the understanding of ADHD and provide insights into distinct biomarkers that can provide guidance for the development of specialized subtype treatments for ADHD. Moreover, the proposed classifiers can help identify the disorder.