Abstract / Summary
Machine learning offers a promising framework for transforming neuroimaging into predictive tools for psychiatric phenotypes. However, a major translational challenge is whether models trained on large, heterogeneous population cohorts can generalize to clinical populations without dataset-specific tuning. Here, we tested the generalizability of neuroimaging-based cognitive prediction models to children with attention-deficit/hyperactivity disorder (ADHD). Using baseline data from the Adolescent Brain Cognitive Development (ABCD) Study (n = 11,747), we trained prediction models across 81 MRI-derived feature sets spanning task-based fMRI contrasts, functional connectivity, structural MRI, and diffusion measures and their multimodal stacking combinations. We tested performance separately in non-ADHD participants and across increasingly stringent ADHD tiers, and further evaluated external generalization in an independent ADHD cohort (Lytle et al., 2020; n=79) without retraining. Internally, the stacked model integrating all feature sets achieved the highest predictive accuracy (r=.57) and reproduced observed group-level cognitive differences, while maintaining comparable performance across ADHD and non-ADHD groups. Externally, stacked working-memory task models generalized to the independent ADHD cohort (r=.44). Together, these findings provide evidence that population-trained, multimodal neuroimaging models can capture cognition-related neural variation across ADHD and non-ADHD groups, while the external validation provides preliminary evidence of transferability across cohorts and measurement contexts.