Optimal Individualized Early Warning for Inpatient Adverse Events

• Activity Number: 311 - Overcoming Practical Challenges in the Design and Analysis of Medical Studies Using Electronic Health Records
• Type: Invited
• Date/Time: Tuesday, August 1, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Suchi Saria* and Hossein Soleimani
• Companies: Johns Hopkins University
• Keywords: Dynamical Prediction with Joint Models ; Scalable Joint modeling ; Survival Analysis ; Electronic Health Records ; Causal Inference from Longitudinal Time Series ; Adverse Event Prediction

Abstract

Inpatient adverse events such as sepsis and respiratory are expensive and preventable. In this talk, we discuss a computational framework leveraging electronic health record data to provide decision support. Our framework jointly models the longitudinal and time-to-event data. To make inference tractable, we propose an embarrassingly parallel stochastic variational inference algorithm. We will show significant improvements over state-of-the-art on accuracy and reliability of the early warning system and discuss takeaways from deploying a preliminary version of this work in the inpatient setting.

In summary, the three key points emphasized in this talk will be: 1) a scalable framework for joint modeling of longitudinal and time-to-event data, 2) introduction to the problem of surveillance for inpatient adverse event, and 3) numerical results comparing proposed framework on a large scale, challenging inpatient dataset.