Predicting Cognitive Functioning in ADHD Using Population-Based MRI Across Large and Small Samples

ObjectiveTo assess whether cognitive prediction models trained on multimodal neuroimaging from a population-based cohort generalize to children with and without ADHD across internal and external datasets. MethodsThis cross-sectional study used task-based and resting-state fMRI, structural MRI, and diffusion tensor imaging from the Adolescent Brain Cognitive Development (ABCD) Study (n = 11,747; mean age = 9.5 years) to train models predicting cognitive functioning. ADHD diagnoses were stratified into four tiers (n = 1,034 to 61), with the remaining participants classified as non-ADHD (n = 10,713). Models were trained using either single neuroimaging feature sets (e.g., task-based fMRI contrasts or cortical thickness) or combined feature sets via stacking. External generalizability was tested using an independent ADHD dataset (Lytle et al.; ADHD, n = 35; non-ADHD, n = 44; mean age = 9.0 years). ResultsIn ABCD, the stacked model integrating 81 neuroimaging types achieved comparable predictive performance for non-ADHD (r =.57) and ADHD (r =.51-.56 across tiers) groups. Key contributors included (a) fMRI contrasts from the NBack task, particularly in the ventral occipital cortex and anterior cingulate cortex (ACC), and (b) task-based functional connectivity from the dorsal attention and posterior multimodal networks. In external validation, combining different fMRI contrasts from the NBack task yielded similar performance for non-ADHD (r =.36) and ADHD (r =.42) participants. ConclusionsPopulation-trained neuroimaging models generalized well to both ADHD and non-ADHD children, underscoring the translational potential of multimodal brain-based models for predicting cognitive functioning in clinical populations.