- Psych 1924
Title: Statistical Machine Learning and Computational Medicine
The digitization of clinical data has created a complex, high-dimensional, and multi-modal summary of a patient's health. The goal of computational medicine is to use these data to generate clinically relevant insights that lead to improved health outcomes. Modern machine learning methods are adept at finding predictive signal in high-dimensional data but not necessarily at revealing the physiology underlying these predictions. In the first part of this talk, I will discuss recent work centered on electrocardiogram (EKG) data and their use for prediction, risk stratification, and patient representation. I will describe a model that predicts adverse cardiovascular outcomes directly from a patient's EKG and compare it to models based on more traditional risk factors. I will then describe a new technique that uses a flexible generative model to visualize and explore variation in the data that is predictive of an outcome of interest. These complex statistical models for high-dimensional data lead to challenging inference problems. In the second part, I will discuss recent work on improving a general class of approximate inference methods. These methodological developments focus on increasing the expressiveness of variational inference approximations and the robustness of the variational optimization procedure.
Andy Miller is a Postdoctoral Research Scientist at the Data Science Institute at Columbia University. He develops probabilistic modeling techniques and inference methods for complex, high-dimensional data in applied areas ranging from astronomy to sports analytics to health care. At Columbia, he works with John Cunningham (Statistics) and Dave Blei (Statistics and Computer Science). Prior to Columbia, he earned his PhD at Harvard University in Computer Science working with Ryan Adams.