The Problem
Every year individual emergency departments (EDs) across the United States see and treat tens of thousands of patients. While the vast majority of those patient visits result in the patient being discharged, the ED continues to be the primary source for admissions to the hospital. [1] Many institutions are seeing ED and hospital volumes continue to rise and historical data are fairly reliable in predicting admission rates when considered over time. While this may be reliable on the macro level, this is not the case on the individual patient level. There have been some attempts to develop a scoring system that can be used at the time of triage to help predict the likelihood that an individual patient will be admitted versus discharged. Unfortunately, the reliability of those systems is poor.[1,2] Moreover, while there is some correlation in initial triage and final disposition it has been found that it is not predictive. [3] This is because many triage systems, including the emergency severity index (ESI), are developed to estimate resource needs per individual visit and not necessarily true acuity. In fact, a recent paper utilizing the NHAMCS database shows that those with the middle acuity (ESI of three) actually have a high resource utilization similar to those within ESI one or two but with admission rates that are far less than those patient populations. [3]
Given that this issue is fairly universal, many EDs have begun to use advanced techniques including computational modeling to predict final disposition status on patients early in their ED visit, sometimes even at triage. While a number of these predictive models show some promise, they are based on a binary output of “admission” or “no admission” [1,2,4,5] Others have attempted to predict one-year mortality following hospital admission, though were not designed to predict in-hospital mortality on the index admission. It is not clear if any models have successfully been able to predict resource utilization (e.g. diagnostics, consultation, staff time needs, etc) while within the ED, length of stay in the hospital after admission, likelihood of escalation of care (RRT or unexpected ICU admission), or in-hospital mortality.[6,7,8] Having such a predictive tool would allow improved patient care by improving patient flow while within the ED by expediting patient assignment to the appropriate care area within the ED. It could also impact patient flow for admitted patients throughout their inpatient stay until their discharge home. [9] Moreover, the value of such a tool to a hospital’s bed control department would be significant as it would allow earlier notification of not only the need for a bed but the specific type of bed an individual patient will need even prior to the completion of the ED evaluation.
Our Solution
Machine learning is becoming a valuable tool within healthcare. By using large volumes of historical patient data models can be built to help predict various clinical outcomes, though typically only in a binary fashion e.g. “admission” versus “no admission”, ‘septic” versus “not septic”, ‘in need of blood transfusion” versus “no need for transfusion”. It is the goal of this project to push the boundaries of a machine learning models to provide a useful array of information regarding individual patient clinical courses that can be used to improve efficiency and patient flow through both the ED and throughout the hospital. This model would lead to develop of a real-time clinical dashboard that could be used by designated staff within the ED to help place patients in treatment locations based on resource needs (e.g. match the needs of the patient with the resource allocation within various areas in the ED) while also taking into consideration likelihood of admission and clinical trajectory upon admission.
By improving the efficiency and timeliness of the clinical care provided to every patient within the ED, regardless of triage leveling or final disposition, there would be an anticipated improvement in ED length of stay without the degradation of clinical care while hopefully improving the patient experience. Additionally, by providing the hospital bed control accurate real-time information around volume of patients who need to be admitted as well as their projected clinical trajectory and length of stay decisions around unit staffing bed management and transfers can be made more accurately.
Impact
This project would transform how DUH manages inpatient admission flow from the ED and could potentially revolutionize how EDs globally approach evaluation and management of all of their patients. By being able to reliably predict which patients are going to need inpatient admission as well as resource utilization while in the ED, nursing and providers can more appropriately manage flow of individual patients through their ED course. In also predicting necessary in-patient level of care and length of stay early in the patient’s ED visit, this model could aid in efficiently moving patients from arrival at the ED’s front doors to placement in the appropriate inpatient bed or area for discharge in the ED in a safer and more timely fashion.
Fenn A. (2020, February 22). Development of Machine Learning Models to Predict Admission from ED to Inpatient and Intensive Units [Poster Presentation]. 2020 Society of Academic Emergency Medicine (SAEM) Southeastern Regional Conference, Greenville, SC. https://ghscme.ethosce.com/courses/2020SAEM YouTube: Development and Validation of Machine Learning Models to Predict Admission….
