Bayesian Transfer Learning for Enhancing Brain-Behaviour Predictions in Small fMRI Samples

• Applicant: Dr. Carsten Gießing
• Subject Area: Biological Psychology and Cognitive Neuroscience ; General, Cognitive and Mathematical Psychology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 564439360

Can brain models of attention, estimated from large datasets, be transferred to smaller samples with different tasks and contexts? Can these models enhance the analysis of smaller datasets? This project explores these questions by employing Bayesian methods for knowledge transfer to examine the transferability of existing models and to develop a new, integrated model with improved generalizability. Particularly in fMRI research, where large sample sizes are often difficult to achieve, Bayesian transfer learning with power priors offers an elegant solution: it enables the integration of prior knowledge from larger datasets into the analysis of smaller samples while simultaneously accounting for the heterogeneity of the data. However, key questions remain unanswered: How do existing brain models of attention differ in their ability to transfer to new contexts? Can an integrated model that combines different approaches further improve transferability? Additionally, the fMRI literature lacks practical tools to provide researchers with easy access to suitable priors for their own analyses. This project addresses these challenges through three closely linked work packages. First, various brain models of attention - based on functional connectivity, network topology, as well as dynamic and static approaches - are systematically compared and integrated into a robust, transferable model. In the second work package, the focus shifts to validating the transferability of these models. Using an independent dataset and power priors, the project investigates how prior knowledge can be optimally transferred to new contexts. A key milestone here is the development of a software tool that enables researchers to use power priors for their studies. Finally, the third work package demonstrates the application of knowledge transfer in a pharmacological fMRI dataset, showing how power priors can significantly enhance the analysis of small sample sizes. By doing so, the project makes a significant contribution to improving the reproducibility and precision of brain-behavior models, paving the way for their application in personalized medicine.

• DFG Programme: Research Grants