DEEP LEARNING FOR FUNCTIONAL CONNECTIVITY PREDICTION OF NON-BRAIN-IMAGING PHENOTYPES

Abstract

Machine learning is an important tool powering many aspects of neuroscience research. A growing literature applies classical machine learning algorithms and deep neural networks (DNNs) to a variety of brain imaging applications. In this thesis, we compared kernel ridge regression with three types of DNN architectures in RSFC-based phenotypic prediction. This is one of the largest empirical evaluations of DNN’s utility in RSFC-based fingerprinting. We showed that kernel ridge regression and three DNNs achieved similar performance for RSFC-based prediction of a wide range of phenotypes. Then, we proposed a novel and simple approach, meta-matching, which leverages the power of large datasets to boost the RSFC-based prediction of unseen non-brain-imaging phenotypes (nBIPs) in small-scale studies. Meta-matching can be utilized in conjunction with any machine learning algorithm. We showed that meta-matching dramatically improved RSFC-based phenotypic prediction on new unseen phenotypes with small sample sizes.