Predicting Task Functional Localizers Using Naturalistic fMRI

Abstract

Functional magnetic resonance imaging (fMRI) data collected during naturalistic stimuli has shown promise for predicting individual traits, biomarkers of disease and functional brain localizations, potentially offering advantages over traditional resting-state approaches. This study investigated the use of interpretable deep learning models to predict demographics and functional task localizer activations from fMRI time-series data collected while participants viewed naturalistic stimuli. Using the data of 143 subjects from the Human Connectome Project, I analyzed 7T fMRI scans from participants watching movies to predict sex, age, and functional localizer activations across multiple cognitive tasks. I employed state-of-the-art machine learning architectures, including DICE and Glacier models, specifically chosen for their interpretable design features that build directed connectivity matrices and produce weighted temporal attention maps. These models aimed to capture dynamic brain activity patterns while maintaining the ability to understand which temporal features drive predictions. The results successfully reproduced previous findings for sex classification but showed poor performance for age prediction, with correlations ranging from -0.175 to 0.243. For functional localizer predictions, models initially appeared to achieve high performance with some specific contrasts having correlations around 0.9 and Dice scores generally above 0.6. However, detailed analysis revealed that these models were primarily predicting group averages rather than learning meaningful inter-subject variability, as evidenced by chance-level subject identification accuracy. This finding contrasts with previous works that demonstrated successful prediction of individual differences in functional localizations. The failure to capture inter-subject variability represents a significant limitation, as individual differences in functional regions of interest are crucial for applications such as pre-surgical mapping and disease prediction. My findings suggest that predicting from raw fMRI time-series may require different approaches than those used here, with preprocessed functional connectivity matrices showing promising results, and highlight the importance of sufficient training data to separate signal from noise when learning directly from naturalistic stimuli. Despite these challenges, this work establishes important methodological foundations and identifies key limitations that must be addressed in future research combining naturalistic stimuli with machine learning for fMRI prediction tasks. The findings emphasize the need for models that can capture individual functional differences while maintaining the interpretability necessary for understanding how naturalistic stimuli drive brain-based predictions.