Prediction of sleep quality scores using dynamic functional network connectivity of young adults: A reproducibility analysis

Abstract

AbstractBackgroundDynamic functional network connectivity (dFNC) estimated from resting‐state functional magnetic resonance imaging (rs‐fMRI) is a potentially powerful approach to investigate human behavior and cognition. However, previous studies barely investigate the reproducibility of dFNC features links with cognition. This study aims to examine whether the data collected in four different sessions reproduce the same results using the sleep scores by applying a dFNC method to resting‐state fMRI (rs‐fMRI) data from the Human Connectome Project (HCP) Young Adult.MethodWe used the Pittsburgh Sleep Quality Index (PSQI) scores and four sessions of rs‐fMRI (15 minutes each) from 833 young adults (age between 22 to 37yrs). We calculated the dFNC for each participant using a fully automated Neuromark independent component analysis and a sliding window technique to estimate dFNCs. A k‐means clustering approach partitioned dFNCs into two group‐level distinct states and subject‐level state vectors. Next, we estimated dFNC features from each individual’s state vector and dFNCs. Finally, we trained a 10‐fold support vector regression model to predict the sleep scores based on the dFNC features. We applied all the procedures mentioned above to all four sessions separately.ResultWe found dFNC features can successfully predict each participant’s night sleep time. The correlation between the actual score and the predicted one was R = 0.0853 (p = 0.0182) for session1, R = 0.0696 (p = 0.0485) for session2, R = 0.0764 (p = 0.0393) for session2, and r = 0.0788 (p = 0.0285) for session4. The dFNC features for none of four sessions could predict other sleep scores.ConclusionWe showed that the dFNC feature could predict the amount of sleep during the night in the HCP Young Adult population. Additionally, we showed the prediction result replicates across all four rs‐fMRI sessions. Future studies are needed to explore the reproducibility of the result within the age range of the HCP dataset.