Deep learning for health outcome prediction

Abstract

Modern medical data contains rich information that allows us to make new types of inferences to predict health outcomes. However, the complexity of modern medical data has rendered many classical analysis approaches insufficient. Machine learning with deep neural networks enables computational models to process raw data and learn useful representations with multiple levels of abstraction. In this thesis, I present novel deep learning methods for health outcome prediction from brain MRI and genomic data. I show that a deep neural network can learn a biomarker from structural brain MRI and that this biomarker provides a useful measure for investigating brain and systemic health, can augment neuroradiological research and potentially serve as a decision-support tool in clinical environments. I also develop two tensor methods for deep neural networks: the first, tensor dropout, for improving the robustness of deep neural networks, and the second, Kronecker machines, for combining multiple sources of data to improve prediction accuracy. Finally, I present a novel deep learning method for predicting polygenic risk scores from genome sequences by leveraging both local and global interactions between genetic variants. These contributions demonstrate the benefits of using deep learning for health outcome prediction in both research and clinical settings.