Resting‐state dynamic functional network connectivity predicts cognition in 37,784 participants of UK Biobank

Abstract

AbstractBackgroundAge‐related cognitive decline after 65 is a well‐known phenomenon, but little is known about how brain functional changes are related to cognitive decline. To this end, previous studies explored the link between functional network connectivity (FNC) estimated from resting‐state functional MRI (rs‐fMRI) and cognitive scores in healthy adults. FNC is often assumed to be static over time. However, this assumption runs contrary to the dynamic nature of brain FNC, and dynamic FNC (dFNC) has been recently introduced to overcome this limitation. The current study investigated the relationship between dFNC estimated from 36,263 individuals from UK Biobank and cognition.MethodWe used the resting‐state fMRI (duration: 5min) data of 37,784 (20,157 females) adults’ brains, demographic information (age:64.06± 7.51), and cognitive scores from the UK Biobank in which the cognitive scores include fluid intelligence (FI), reaction time (RT), and pairs matching (Pairs). We adapted group independent component analysis to extract 53 data‐driven components for the whole brain using a fully automated approach (Fig.1: Step1). Next, we used the sliding window and Pearson correlation to estimate the dFNC among 53 components (Fig.1: Step2). We used k‐means clustering to put dFNCs of all individuals into three separated states and calculated individual state vectors (Fig.1: Step3). Then, we estimated 32 dFNC features based on three estimated states and state vectors (Fig.1: Step4). Finally, we trained a two‐fold cross‐validation support vector regression to predict the cognitive scores (Fig.1: Step5).ResultThe estimated dFNC feature predicted FI score with high accuracy (the correlation between measure and predicted score is R = 0.043, p = 6.6e−17, Fig.2A), mean time to correctly identify match in RT task (R = 0.065, p = 1.9−34, Fig.2B), and time to complete Pairs task (R = 0.049, p = 1.3e−20, Fig.2C).ConclusionHere we explored the link between dFNC features and cognition in the most extensive dFNC study and found that estimated dFNC features can successfully predict cognitive scores in the UK Biobank dataset. Even though the correlation values are very low, the results are still very significant due to large N. Future study is needed to explore the difference between the subgroup with high versus low prediction accuracy.