Meta-Learning for Generalizable Connectome Modeling Across Heterogeneous Atlas Spaces

Leveraging large-scale neuroimaging datasets to improve predictive modeling in smaller, specialized cohorts is essential for translating research findings into clinical applications. To address this challenge, we propose a meta-learning inspired approach that enhances phenotype prediction by exploiting inter-phenotype correlations across independent datasets, without requiring uniform imaging data processing. By employing mappings derived from a small, independent cohort with data processed using different atlases, we align the data in the meta-training set to the target space of the meta-testing set. This approach eliminates the need for extensive data reprocessing. Evaluation in the UK Biobank dataset shows that our framework significantly improves predictive performance on unseen phenotypes, demonstrating its potential to generalize connectome-based models to small, heterogeneous cohorts.