A hospital’s emergency department swells with patients, a chunk of whom will need to be admitted to the facility’s inpatient unit.
But decision-makers in emergency care face a dilemma: When deciding to admit a patient, they need to make sure there will also be enough hospital beds for patients coming out of surgery, as well as those transitioning out of intensive care units. To make any admission decisions, they first need to know not just how many beds are currently available, but also how many beds will be needed and how many patients will be discharged to open up beds in the near future.
A team of University of Wisconsin-Madison industrial engineers determined a way to help forecast discharges, almost in real time. In collaboration with UW Health and the Department of Emergency Medicine, the researchers have developed a mathematical model that allows for more granular predictions—covering the next hour or next four hours, rather than full-day projections that are typically used in the healthcare industry.
The researchers detail their model in a paper published in the Journal of Medical Internet Research.
Industrial and systems engineering PhD student Fernando Acosta-Pérez and Associate Professor Gabriel Zayas-cabán say their work could allow hospital decision-makers to work more nimbly and cut down on emergency department “boarding,” in which patients are admitted but left waiting in the emergency department for open beds. Studies have shown that boarding is not only bad for the hospital system but also leads to worse patient health outcomes.
“It’s very hard to effectively coordinate and determine how to route patients,” says Acosta-Pérez, a third-year PhD student from Puerto Rico who is the lead author on the paper. “What we’re doing a little differently is we have constant predictions every hour of the day. So we can definitely use the information in a better way, rather than just one single prediction for the entire day. When you need to make decisions now, you need an updated forecast with the most recent information.”
Acosta-Pérez used historical data from UW Health to build and train the model, which uses a collection of algorithms called decision trees to generate its predictions. In partnership with the UW Health data science team, the researchers are continuing to test their model to further validate and refine it. To make that possible, UW Health data scientists created a pipeline to generate real-time data.
In addition to his PhD research, Acosta-Pérez is also interning with the UW Health data science team.
“Fernando not only had access to a richer data set, but he was interacting with these practitioners on a regular basis for key context and experience,” says Zayas-Cabán, who’s a longtime collaborator with Brian Patterson, an associate professor of emergency medicine and senior author on the paper.
For his dissertation research, Acosta-Pérez plans to employ additional mathematical techniques to develop more advanced models that can not only offer predictions but also provide prescriptive recommendations across a network of affiliated hospitals.
Acosta-Pérez, who earned a National Science Foundation Graduate Research Fellowship, says the chance to pursue research while also gaining practical experience with UW Health was one of the reasons he chose UW-Madison for his graduate studies.
“I have this opportunity to not only do research,” he says, “but also hopefully have some meaningful impact and use the stuff that I do for improving systems, which is what industrial engineers try to do.”
Other authors on the paper include Sabrina Adelaine, a program director with UW Health; Frank Liao, a clinical adjunct assistant professor in the Department of Emergency Medicine and senior director of digital health and emerging technologies at UW Health; and Justin Boutilier, a former assistant professor of industrial and systems engineering now at the University of Ottawa. Brian Patterson is also administrative director for clinical AI at UW Health and medical director for predictive analytics on the UW Health informatics teams.
Top photo caption: PhD student Fernando Acosta Pérez uses machine learning and operations research to improve healthcare systems. Photo: Joel Hallberg