Comparing Deep Learning Models for Age Prediction Based on the Resting State fMRI Dataset from the "Brain Dock" Service in Japan

In Japan, many hospitals provide the unequaled service of medical check-up called the "Brain Dock"; however, there is a paucity of studies aimed at leveraging functional magnetic resonance imaging (fMRI) in hospitals. We obtained the resting-state fMRI (rs-fMRI) scans of about 695 patients who accessed this service for preventive medicine against Alzheimers Disease. In this study, we created deep learning models for age prediction with the ultimate aim of arriving at standard protocols for introducing rs-fMRI into clinical settings, particularly in Brain Docks. With that view, an assemblage of modeling conditions was attempted, changing multiple parameter values, features based on data extraction methods (region of interest-wise mean blood-oxygen-level-dependent lump-sum time series models or dynamic functional connectivity models), deep learning algorithms (Transformer, Multi-task Transformer, and unidirectional or bidirectional long short-term memory models), and different atlas-dependent brain region segmentation methods (including the Automated Anatomical Labeling and Harvard-Oxford atlases). As a result, a robust and highly significant correlation was obtained between actual and predicted ages from all types of methodologies. In addition, we determined that some conditions had a relatively large impact on prediction performance based on extended comparisons. The accuracy decreased, particularly according to the choice of atlases but with the same modeling conditions. Notwithstanding, we found that atlases based on intrinsic functional connectivity provided significant prediction accuracy even with a small number of regions to a similar extent as networks lowered spatial granularity. Moreover, we found that multi-task learning with other phenotype data (related to gender differences) was possible, but did not improve the prediction accuracy as much as expected. Despite these limitations, our results could provide a hopeful prospect of introducing fMRI into the field of neuro-clinical practice. HighlightsO_LIAge prediction modeling was performed using the rs-fMRI data of a Japanese "Brain Dock" service targeting mainly elderly people, which made prediction more challenging. C_LIO_LIA robust and highly significant correlation was obtained from all types of methodologies changing data extraction methods, deep learning algorithms, and brain atlases. C_LIO_LIAge prediction was successful even with a very small number of regions exclusively based on intrinsic functional connectivity networks, although there was a difference in the achievement of modeling. C_LI