Predicting phenotypic measures from the functional connectome with geometric deep learning
Approved Research ID: 58302
Approval date: August 13th 2020
The overarching goal of this project is to use an advanced form of deep learning - geometric deep learning - to predict the traits or behaviours of an individual from brain imaging data. Previous research has shown that the activity of the brain at rest (that is, in the absence of any specific task) is unique, acting as a sort of "fingerprint" for each individual. Moreover, it has been shown that these "fingerprints" can be used to predict certain aspects of human behaviour as well as disease states at the individual level. However, current approaches for predicting the relationship between this brain activity and behavioural measures first simplify the "fingerprint" by disregarding its inherent graph structure. Geometric deep learning has recently emerged as a promising approach for building predictive models able to actually exploit the structural features of graphs and is hence well-poised to provide a better tool for this endeavour. It is our aim, therefore, to use geometric deep learning to develop a neural network capable of learning the complex relationship between an individual's brain activity at rest and their behaviour.
This project will take place over the course of the next 3 years, as part of a PhD thesis on applications of geometric deep learning for neuroscience. By providing the scientific community insight into the utility of geometric deep learning for building predictive models using brain imaging data, it will facilitate the development of research-related and clinically-oriented diagnostic tools capable of working directly with these graph-based measures.