Predicting Future Cognitive Development Using Past Brain FMRI with Spatiotemporal Graph Transformers Guided by Prior Information

Predicting individualized cognitive developmental outcomes in early infancy using past (earlier) neuroimaging data has always been a hot topic. Limited progress has been made due to a lack of methods capturing highly complex predictive spatiotemporal changes in infant brains. Instead of using cross-sectional anatomical information, we propose using existing (i.e., past) longitudinal brain functional connectome that is considered more informative for individualized developmental trajectory prediction. Leveraging a pre-trained large language model, we embody a combination of all the past and the target age time points, along with the spatial and functional information of all brain regions, as prior knowledge, to guide multi-task spatiotemporal graph transformers focusing on age-sensitive developmental connectomic dynamics, thus facilitating accurate developmental curve prediction. Our method, coined as Spatiotemporal Transformer with Information Guidance (STRING), achieved the best performance on a large dataset compared with other state-of-the-art methods. We also identified key brain regions with interpretability of STRING.