Brain-wide association studies (BWASs) have attempted to relate cognitive abilities with brain phenotypes, but have been challenged by issues such as predictability, test-retest reliability, and cross-cohort generalizability. To tackle these challenges, we proposed a machine learning "stacking" approach that draws information from whole-brain MRI across different modalities, from task-functional MRI (fMRI) contrasts and functional connectivity during tasks and rest to structural measures, into one prediction model. We benchmarked the benefits of stacking using the Human Connectome Projects: Young Adults (n = 873, 22-35 years old) and Human Connectome Projects-Aging (n = 504, 35-100 years old) and the Dunedin Multidisciplinary Health and Development Study (Dunedin Study, n = 754, 45 years old). For predictability, stacked models led to out-of-sample r∼0.5-0.6 when predicting cognitive abilities at the time of scanning, primarily driven by task-fMRI contrasts. Notably, using the Dunedin Study, we were able to predict participants' cognitive abilities at ages 7, 9, and 11 years using their multimodal MRI at age 45 years, with an out-of-sample r of 0.52. For test-retest reliability, stacked models reached an excellent level of reliability (interclass correlation > 0.75), even when we stacked only task-fMRI contrasts together. For generalizability, a stacked model with nontask MRI built from one dataset significantly predicted cognitive abilities in other datasets. Altogether, stacking is a viable approach to undertake the three challenges of BWAS for cognitive abilities.
Keywords: cognitive abilities; generalizability; reliability; stacking; task fMRI.
© The Author(s) 2025. Published by Oxford University Press on behalf of National Academy of Sciences.