UMD and Johns Hopkins develop model to predict hospital bed count

Sean Barnes, an assistant professor at the business school, helped design a computer program which aims to predict hospital patient discharges.

When the demand for hospital beds exceeds the supply, doctors might have to turn patients away. But a computer model developed by University of Maryland and Johns Hopkins University researchers might help doctors better determine how many patients will be discharged and how many beds will be available on any given day.

“Hospitals tend to be highly utilized,” said Sean Barnes, an operations management professor involved with the model, “which means it can be hard to bring in new patients.”

Barnes and other researchers designed a computer model that could help solve this problem. A paper detailing their research was published online in the American Medical Informatics Association journal in August.

Many hospitals use a management technique called real-time demand capacity, Barnes said. With this process, a team of clinicians each morning try to evaluate and predict which patients are likely to go home by the afternoon — about 2 p.m. — and by the end of the day. This helps them estimate how many new patients they can expect to see that day.

The researchers’ technique automates that process.

“We tried to build a model that could match the predictions of the clinician teams,” Barnes said. “We wanted to see if we could create a computer model to try and predict when patients are likely to leave and to predict when beds will be free and doctors can bring in new patients.”

READ MORE: University of Maryland professor to study service dog therapy for PTSD

Their model takes simple information available from hospital systems at 7 a.m. each day, Barnes said, such as age, gender, race, length of stay, presence of certain symptoms and even what day of the week it is, to try to replicate clinicians’ predictions. They tested the computer model by comparing its predictions side-by-side with those of the real-time demand capacity clinician teams, using about nine months of data.

“The model exceeded the accuracy of the clinicians in almost all cases,” Barnes said.

On an individual patient basis, the model matched the prediction performance of clinicians well. And researchers found the model was more accurate in predicting the overall number of daily discharges.

Barnes said researchers designed and trained the model for about two years before the nine-month testing. They evaluated the model’s performance on close to 9,000 patient stays overall.

“We try to build tools that support clinicians’ decision-making,” said Scott Levin, an associate professor in the department of emergency medicine and civil engineering at Johns Hopkins University involved with the model. “Clinicians can make these decisions, but it takes time and coordination … and this technology enables hospitals to do this process more automatically.”

Levin said this model could help doctors save time so they can focus on more important hospital duties. The model isn’t designed to replace clinicians’ evaluations, but in combination, hopefully the two methods can improve overall efficiency and effectiveness, he said.

“We would like to reserve time for the clinicians to focus more on patient care and not have to spend as much time thinking about operational issues,” Barnes said. “We are hoping this model can be an enhancement, to save them some time that they can dedicate to direct care of the patients.”

Margrét Bjarnadóttir, a management science and statistics professor who was not involved with this research, said analytics and data have proven useful in the health care industry. This research is an example of how applying analytical models in health care can improve efficiency and decision-making, she said.

READ MORE: UMD study on Twitter looks into when people retweet the most

“One interesting thing that makes this model stand out is [the researchers] compare the model to the intuition of experts,” Bjarnadóttir said. “There is a lot of pressure in healthcare for delivering efficient, high-quality patient care and this model is an example of how this can be done.”

Barnes said the next step in this research is to broaden the scope of how the model can be applied.

Levin said researchers are hoping to test the model in different hospitals and on different units and continue to improve the technology.

“Now what we want to do is explore the performance of it,” Barnes said. “Not just on one medical unit, but we want to see how generalizable it is in other environments.”