Accurately predicting which Emergency Department (ED) patients are at high risk of leaving without being seen (LWBS) could enable targeted interventions aimed at reducing LWBS rates. Machine Learning (ML) models that dynamically update these risk predictions as patients experience more time waiting were developed and validated, in order to improve the prediction accuracy and correctly identify more patients who LWBS.

The study was deemed quality improvement by the institutional review board, and collected all patient visits to the ED of a large academic medical campus over 24 months. The first 18 months of data were used to develop two types of classification models using XGboost: (1) a static model that uses patient and ED census information at the time of arrival to predict the risk to LWBS; and (2) a dynamic model that updates the predictions based on new information after 30 min for patients who are still waiting in the ED. The final six months of data were then used as an experimental period to test the accuracy of the two models.

For a cohort of 150,959 ED patient arrivals, the mean patient age was 46, 51 % of arrivals were female, and 2.17 % of patients LWBS during their wait. The models achieved an area under the receiver operating characteristic curve (AUROC) of 0.86 when dynamically updating LWBS risk levels over time. This was in contrast to an AUROC of 0.80 for the static model from past literature that does not update the score after the moment of arrival. Over the experimental period, the dynamic model also showed the ability to reduce the number of missed LWBS cases by approximately 50 % as compared to the static model, without incurring any additional false-positives.

Dynamically updating patients' risk to LWBS as their wait goes on significantly reduces the number of patients who LWBS that are missed by prediction models as compared to traditional static prediction approaches.