Abstract:

Electronic Health Records (EHR) data constitute a relatively new data source that contain a running tally of a patient’s clinical changes. As such, they are an appealing resource for clinical analysis, particularly risk predic- tion. While these data are potentially powerful, they inherently have a number of challenges such as many potential predictor variables, sparse and irregular measurements over time, and data that may be informatively not observed. As as result, developing robust risk models can be challenging. Using data from our institution’s EHR system, we illustrate the various considerations necessary for developing a dynamic risk score for inpatient deterioration. We choose a computationally efficient time varying Cox model and show how the model can be adapted to incorporate different data complexities. We compare our results to an Early Warning Score currently implemented in the EHR sys- tem, showing that fitting a model with one’s own institution’s data results in better performance, even when using the same predictor variables.

[pdf]

Advisor: Dr. Ben Goldstein, Dr. Beka Steorts