Overcoming Vulnerabilities of Machine Learning in Neuroimaging Applications

Neuroimaging has been widely used to non-invasively probe the structural and functional changes in many neurological and psychiatric diseases. Techniques such as functional magnetic resonance imaging (fMRI) generate a wealth of dense and high-dimensional measurements of the brain. Through machine learning, a powerful class of statistical methods, such measurements can be used to make personalized predictions that inform clinical care, such as diagnoses, prognoses, or treatment outcome trajectories. However, applications of machine learning must meet two critical requirements: 1) there must be sufficient training data and 2) data samples should be independent and identically distributed (iid). In this dissertation, I propose two novel methods to overcome these vulnerabilities in neuroimaging and other biomedical data, and I develop two promising applications of machine learning and neuroimaging to address critical needs in neurological and psychiatric disease. First, I demonstrate how standard machine learning approaches can be applied to predict future Parkinson's Disease severity from resting-state fMRI. By generating personalized prognoses for this debilitating neurodegenerative disease, better care decisions can be made. Next, I address the requirement for sufficient training data in machine learning, without which models tend to overfit and generalize poorly to new data. I propose a method to perform data augmentation, where additional data is simulated by transforming existing data, for 4D fMRI timeseries. Using this augmentation method, I develop predictors of antidepressant response from pre-treatment task-based fMRI. This targets a pressing need for individualized treatment selection tools in depression care, where it currently takes weeks to months to test various medications until adequate relief is found. Finally, I propose a method to construct neural networks, a potent class of machine learning models, that can specially handle non-iid data. This is of broad importance in biomedical data, which is frequently non-iid with samples clustered by experimental batch, study site, etc. Inadequate handling of cluster effects leads to confounding biases and poor generalization in conventional models. My method, tested across several biomedical applications, improves both performance and generalization while affording greater interpretability of cluster effects in the data.