A multimodal machine learning framework for ADHD diagnosis and gender classification using neurobehavioral data

Attention-deficit/hyperactivity disorder (ADHD) is a complex and prevalent neurodevelopmental disorder marked by inattention, impulsivity, and hyperactivity, which collectively impair daily functioning. In response to the demand for scalable and objective diagnostic tools, this study presents a machine learning (ML) framework that integrates demographic, behavioral, and neuroimaging-derived data to both classify ADHD and infer gender distributions. Four data modalities were utilized, including clinical metadata, fMRI-based functional connectome matrices, behavioral assessments, and diagnostic labels. For ADHD classification, Random Forest was used for feature selection, followed by XGBoost for modeling, achieving an accuracy of 81.02%, a macro-averaged Fl-score of 0.76, and a recall of 92% for ADHD-positive individuals—demonstrating strong screening potential. For sex prediction, LightGBM was used for feature selection, followed by SVM for modeling with threshold optimization and SMOTE balancing produced an overall accuracy of 87%, with balanced performance across both male and female classes (macro F 1 -score of 0.87). These results underscore the clinical relevance and generalizability of the proposed model, while highlighting the potential of behavioral and connectome features in data-driven ADHD diagnostics.