Tue, Aug 14, 2018 @ 10:00 AM - 12:00 PM
Title: Learning to Diagnose from Electronic Health Records Data
Ph.D. Candidate: David C. Kale
Date and Time: Tuesday, August 14, 2018 at 10:00 AM in GFS 108
Committee: Greg Ver Steeg (Chair), Aram Galstyan, Gaurav Sukhatme, and Raghu Raghavendra
With the widespread adoption of electronic health records (EHRs), US hospitals now digitally record millions of patient encounters each year. At the same time, we have seen high-profile successes by machine learning, including superhuman performance in complex games. These factors have driven speculation that similar breakthroughs in healthcare are just around the corner, but there are major obstacles to replicating these successes. In this talk, we will discuss solutions to some of these challenges in the context of learning to diagnose, which involves building software to recognize diseases based on the analysis of historical data rather than expert knowledge. Our central hypothesis is that we can build such systems while minimizing the burden of effort on clinical experts. We will present results from one of the first successful applications of recurrent neural networks to the classification of multivariate clinical time series. We will then show how to extend this framework to model non-random missing values and heterogeneous prediction tasks. Finally, we will describe a public benchmark for clinical prediction and multitask learning that addresses the crisis of reproducibility in clinical machine learning and lowers the barrier to entry for new researchers. We will also spotlight additional research that considers nearest neighbor approaches and weak supervision in the absence of ground truth labels. We conclude by considering the broader impact of information technology on healthcare and how machine learning can help fulfill the vision of a learning healthcare system.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon