MIT Sloan Health Systems Initiative
Getting Treated at the ED as Painlessly as Possible
Nobody goes to the Emergency Department (ED) hoping for an extra long wait to be seen. Rather, long wait times may be considered an immutable characteristic of going to the ED for medical care. Professor Georgia Perakis and her collaborators (Assistant Professor of Emergency Medicine at UMass Medical, Mo Canellas; Professor Dessislava Pachamanova, Babson College, and MIT PhD students Omar Skali Lami and Asterios Tsiourvas) challenged that assumption. Their recent research using Machine Learning tools delved into ED Length of Stay (LOS) data to unearth the significant factors that contributed to long wait times. More than just describing the issue, the team constructed a predictive model that can be adapted to address similar issues in any ED.
Other researchers have considered LOS in their models, but this team came up with the new tactic of dividing LOS into three parts and considering each one as a separate input to their algorithm. LOS Is divided into:
- From the time the patient walks into the ED until they are put into a room in the ED
- From the time the patient is put into a room in the ED until they are discharged
- From the time the patient is admitted until they are brought to an inpatient bed
Dividing LOS into three parts was key to developing the final effective algorithm that could predict length of stay and recommend ways a hospital ED could improve.
To construct and train their model, the researchers used data from eight months of visits to the University of Massachusetts Medical Center Emergency Department collected from adult patients who were seen between July 2019 and February 2020. This resulted in data for 36,597 unique encounters. The research team chose this time period so as not to confound their calculations with data from COVID patients.
The researchers’ algorithm is able to predict how long any individual will need to be at the ED. Most interesting, the model does not need to know the patient’s diagnosis or prognosis. The only relevant input attributes are how the patient appears at the ED. In practical measures, just from the information a patient gives at the triage desk upon arriving at the ED is enough for the model to predict that patient’s LOS. Since patients with different severities of presenting issues are assigned to varying spots in the ED, having the patient’s LOS upon arrival may help assign that person to the best spot in the ED to be seen.
The model could predict length of stay for a specific patient, but when it comes to improving overall LOS in the ED, it turns out that why the patient comes to the ER nor any other health-related characteristic does not matter.
The research results revealed that the most important factors affecting LOS are the ED’s capacity, flow and resources. For example, things like how many and what types of beds are available, what staffing is available. And this model can be adapted for a busy city academic hospital, in a small rural hospital and anything in between.
Hospitals and the ED departments are at least theoretically in control of their capacity, flow and resources. So, if they want to improve the ED length of stay, the main factors that need to be manipulated are all inhouse. The methods for reducing LOS do not depend on who shows up at the ED for treatment
This research made evident both that a patient’s LOS can be predicted upon presenting at the ED and that overall ED LOS has the potential to be reduced by the hospital's strategic decisions. The first point may help mitigate slowdowns and bottlenecks in real time. If it seems like several incoming patients will need extensive help, those resources can be diverted to the ED even before those patients are seen. However, to make lasting changes to reduce LOS overall, hospital leaders must make choices about resources and allocation in house to achieve their LOS goal. The team is currently also working on a prescriptive model that would address the resource allocation issue.