Integrating Structural Brain MRI and Clinical Phenotypes for Automated ADHD Diagnosis: A Multimodal Deep Learning Approach

Attention-Deficit/Hyperactivity Disorder (ADHD) diagnosis is largely based on subjective evaluations of behavior, which highlights the requirement for objective, neurobiologically based biomarkers. Structural magnetic resonance imaging (MRI) has pointed to very delicate neuroanatomical changes related to ADHD; however, the high variability between subjects makes its reliability doubtful if it is only used for diagnostic purposes. Here we introduce a multimodal deep learning model that combines three-dimensional structural MRI and clinical phenotypic variables (such as age, intelligence quotient (IQ), and medication status) for automated ADHD diagnosis to be more accurate. The proposed design includes: (1) a 3D convolutional neural network for the progressive extraction of structural brain features, (2) a multilayer perceptron for the representation of categorical and continuous phenotypic descriptors, and (3) an adaptive gated fusion component that adjusts modality weighting individually and dynamically at the subject level. The system was tested on the Peking 1, subset of the ADHD200 Preprocessed Anatomical Dataset with 5-fold stratified cross-validation and reached an accuracy of 87%, an AUC of 0.88, a sensitivity of 0.88, and a specificity of 0.79. Comparative experiments reveal that the multimodal adaptive strategy is better than the unimodal baselines, which means that subject-specific fusion of neuroimaging and clinical features offers a more complete depiction of the ADHD heterogeneity.