Individual differences in cognitive abilities ultimately derive from differences in brain function. Recent studies have been able to predict, to some degree, individual differences in intelligence from the connectivity pattern among brain regions obtained at rest with functional magnetic resonance imaging (rs-fMRI). However, it remains unclear how reliable these findings are, how well they replicate, and to what extent they generalize to other samples. These questions have also limited our understand of what it is in the neuroimaging data that drives these predictions; for example, are there specific brain regions, or specific ways that the data are processed that make a difference? To address these questions this project will begin with a successful initial finding, predicting intelligence from rs-fMRI with a particular approach, in a large data set (the Human Connectome Project data, HCP). Building from this initial finding, a series of research aims will then investigate how different kinds of analyses might yield different results, how statistically reliable the findings are, how well they replicate, and how they generalize to databases other than the HCP. These findings will be high methodological value to all scientists working in this field and will also yield initial answers to important questions regarding the neural basis of intelligence. All work will use open-science practices including but not limited to pre-registration and data sharing of data and software.
This project capitalizes on recent success in predicting general intelligence (g) from resting-state fMRI (rs-fMRI) data in the Human Connectome Project dataset (HCP). Its principal aims are to investigate the reliability, reproducibility, and generalizability of this finding. A first aim will quantify the effect of brain alignment, rs-fMRI denoising, brain parcellation, and model-learning strategy on the prediction of intelligence from rs-fMRI in the HCP. The aim will quantify how choices at key intersections in this processing decision tree affect final prediction results. This investigation will provide a valuable inventory of possible processing pipelines, and the difference that different parameter choices make; aim to yield a single "best" combination of analytical choices; and explore which anatomical brain regions, and networks, can best predict intelligence. A second aim will add graph-theoretical summary features and externally driven brain states to improve the prediction of intelligence in the HCP. Do features derived from rs- or task-fMRI yield substantially better predictions? Do models built separately from each paradigm, and for different tasks, point to shared anatomical regions? Combining features from both paradigms, what is the best prediction we can obtain? To investigate the generalizability of results obtained, the results will be replicated across three independent datasets: the Enhanced Nathan Kline Institute - Rockland Sample (NKI-RS; target 1000 participants, 6-85 year-olds), the NIH Adolescent Brain Cognitive Development dataset (ABCD; target 10,000 participants, 9-10 year-olds), and the Cambridge Center for Ageing and Neuroscience (Cam-CAN; 700 participants, 18-88 year-olds). These pre-registered studies will quantify how robust are the brain predictors of intelligence (to different subject samples, and different MRI acquisition methods).
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
