Haraldur Hallgrimsson passes his PhD proposal on "Towards Interpretable Models of Health".
Committee: Ambuj Singh (Co-chair), Scott Grafton (Co-chair), Subhash Suri, Xifeng Yan
Abstract: Healthcare is at a digital inflection point. With a total volume of over two thousand exabytes of healthcare data expected to be collected in 2020 at a nearly 50% yearly increase, clinical trials and research are increasingly fueled, and stressed, to become more efficient. Along an axis of the expense required to acquire a health record, in this talk I consider two extremes: inexpensive wearable trackers of physical activity, and high-fidelity medical images. The low cost of acquiring so called digital health signals from wearables enable massive population studies and better patient experience with a focus on preventative care; while medical images enable in-depth study of specific aspects of the human anatomy and diagnosis thereof. Both of these entail challenges in developing interpretable models that add confidence to their decisions.
In this talk I present an overview of my recent and ongoing work in these two application areas. In particular, I will focus on characterizing localized regions of white matter connectivity, as imaged by diffusion-weighted MRI brain scans. These methods build upon a Bayesian deep learning framework to quantify uncertainty of the small-sample and high-dimensionality data set. In particular, the characterization is focused on identifying sparse predictive subsets, and quantifying the wiring complexity and heterogeneity of parts of white matter.