Longitudinal Causal Inference Using EHRs

• Activity Number: 558 - The Big Data Revolution in Health Care: Promise and Potential
• Type: Topic Contributed
• Date/Time: Wednesday, July 31, 2019 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Roy Adams* and Katharine E Henry and Hossein Soleimani and Michael Rosenblum and Suchi Saria
• Companies: Johns Hopkins University and Johns Hopkins University and University of California - San Fransisco and Johns Hopkins Bloomberg School of Public Health and Johns Hopkins University
• Keywords: causal inference; electronic health records; early warning systems

Abstract

There are many causal questions we may be interested in answering in a clinical setting such as: How important is early detection of a condition? How important is medication timing? Or, how will a patient respond to different doses of a medication? Electronic health records (EHRs) present a rich non-experimental source of data for answering these types of questions; however, they also present a large set of challenges that are not present in many other non-experimental causal inference settings. These challenges include high-frequency irregularly sampled measurements, complex observed treatment policies, and potentially unobserved key factors such as a patient's disease state.

In this talk, we will present an approach to evaluating an early warning system by considering the effect of initiating an intervention based on the output of this system. We treat this as a time varying exposure and use longitudinal causal inference methods to adjust for time varying confounders. Finally, we use this method to evaluate the effect of delivering antibiotics based on a sepsis early warning system.