Development, Validation, and Clinical Trial of a Prediction Tool for High-Risk Emergency Department Revisits

Emergency department (ED) visits have grown tremendously in Taiwan over the past few years, reaching approximately 7 million visits in 2019. This growth in ED visits has led to unprecedented levels of ED crowding, resulting in delays in care and increased morbidity and mortality in the ED. One of the complications of ED crowding is unsafe ED discharge leading to return ED visits. About 5-10% of the patients return to the ED within 3 days of the initial ED visit. Return visits are not only burdensome but also costly. Some return ED visits are even potentially dangerous, such as intensive care unit (ICU) admission, receiving emergency surgery or cardiac catheterization, and in-hospital cardiac arrest (IHCA) during the revisit to the ED. These "high-risk" return ED visits are important in that they may reflect suboptimal initial ED care. Early identification of these high-risk events and appropriate interventions are critical for clinicians and patients. The investigators previously estimated that about 0.1% of the ED visits would become these high-risk ED visits.

High-risk revisits are relatively rare events, which require a large sample size for investigation. To this end, the increasing use of electronic health records (EHR) has opened the door to research rare events using big data. With the gradual adoption of EHR in major sections of the NTUH (outpatient in 2006; inpatient in 2007; and emergency department [ED] in 2009), investigators are now able to conduct clinical research using a wealth of data components in the EHR. For example, from 2013 to 2022, there have been >1 million ED visits to the NTUH. The large sample size also enables studies of rare but important events, such as serious adverse events in return ED visits.

In this proposal, the investigators aim to develop and validate a vital-sign trajectory model (GBTM) and a machine learning/artificial intelligence (ML/AI) based prediction model for the high-risk ED revisits. The investigators will develop and validate these predictive tools using the NTUH data from 2013 to 2022, totaling approximately 1 million ED visits. In phase 2 (prospective validation), the investigators will prospectively collect data, fine tune our model, and optimize the system structure. In phase 3 (intervention), the investigators will embed the model into our hospital information system (HIS) in real-time to identify high-risk patients on ED discharge. If a patient is classified as high-risk, the computer user interface will recommend re-evaluating the patient and observing. At the discretion of the treating physician, wearable smart wristbands will be distributed to borderline high-risk patients who are still sent home to augment vital signs monitoring for 7 days. A dedicated case manager will review the records daily and consult the treating physician if red flags occur. A telemedicine session will be initiated if needed. The detection and intervention components together form the response system, the Emergency Department High-risk Revisits Prevention System (ED-HIKRS). In the intervention phase, the investigators will enroll and randomize 10 attending physicians into the ED-HIKRS or the usual care arm for 5 months, with each of them treating 200 patients (patients clustered within physicians). The pragmatic trial will stop for 2 months as a wash-out period, and an interim analysis will be performed. The physicians will then cross over, and each of them will treat 200 patients for another 5 months. A total of 4,000 adult patients will be enrolled. The primary outcome is the rate of high-risk revisits (ICU admission, receiving emergency surgery, and IHCA within 7 days of ED discharge) or same outcomes within 24 hours of hospitalization through the ED revisit. The secondary outcomes include general ED revisit rates, revisit admission rates, and ED mortality during the revisit.

The proposed study is highly significant because it addresses high-risk ED revisits, an understudied yet increasingly important area. The two state-of-the-art models, if successful, can be readily implemented in the EHR system for real-time prediction. The intervention of ED-HIKRS, if successful, may possibly intercept unsafe ED discharges and prevent subsequent serious adverse events. Our long-term goal is to deploy our system in existing clinical workflows to assist clinicians. This model may be generalizable to other health care systems countrywide.