A view-engage-predict framework for enhancing brain-behavior mapping with naturalistic movie-watching fMRI

Most brain-behavior mapping studies rely on resting-state functional connectivity (FC), but this approach has known accuracy limits and can be outperformed by movie-watching FC. Here, we present a novel deep neural network framework to predict cognitive scores and sex from FC during naturalistic movie viewing, and examine how movie content and its ability to synchronize brain activity across individuals relate to prediction performance. We show that FC from movie-watching generally outperforms resting-state FC - even when compared to five times more temporal data - with sensory and higher-order brain networks emerging as the most important for prediction. Using both static and sliding-window dynamic FC approaches, we find that higher cognitive prediction accuracy is positively associated with greater inter-subject synchrony and the duration of human faces and voices in the movies; these effects were not found for sex prediction. This work underscores the promise of naturalistic movie viewing as a powerful tool for probing individual differences in the brain and revealing neural underpinnings of human behavior. Deep learning on movie-watching fMRI unravels the role of inter-subject neural synchrony and social stimulus content in shaping more accurate and interpretable brain-behavior prediction – pointing to a new design principle for naturalistic neuroimaging.