Forecasting mortality associated emergency department crowding

TL;DR

This study tackles the problem of predicting mortality-associated emergency department crowding by defining a crisis as $EDOR>90\%$ and forecasting its occurrence using a LightGBM model trained on anonymous hourly occupancy data from a large Nordic ED. It achieves section-wise predictions with lead times up to 1 p.m., showing AUROC values up to $0.87$ for bedoccupying patients and demonstrating that calendar factors, especially weekday and holidays, strongly influence forecast accuracy. The approach uses an expanding window retraining scheme and SHAP-based feature attribution to reveal key drivers, while maintaining data privacy by avoiding patient-level information. The findings support the feasibility of deploying preemptive interventions based on timely, anonymized forecasts to mitigate mortality risk from crowding and improve ED operations.

Abstract

Emergency department (ED) crowding is a global public health issue that has been repeatedly associated with increased mortality. Predicting future service demand would enable preventative measures aiming to eliminate crowding along with it's detrimental effects. Recent findings in our ED indicate that occupancy ratios exceeding 90% are associated with increased 10-day mortality. In this paper, we aim to predict these crisis periods using retrospective data from a large Nordic ED with a LightGBM model. We provide predictions for the whole ED and individually for it's different operational sections. We demonstrate that afternoon crowding can be predicted at 11 a.m. with an AUC of 0.82 (95% CI 0.78-0.86) and at 8 a.m. with an AUC up to 0.79 (95% CI 0.75-0.83). Consequently we show that forecasting mortality-associated crowding using anonymous administrative data is feasible.

Figures (7)

• Figure 1: Graphical representation of the concept of the early waning software. The red area shows EDOR$\ge$90 % which we have previously associated with increased 10-day mortality Eidsto2023. The black and red lines represent the historical hourly EDOR among bedoccupying patients during January 3, 2018. The vertical dashed line on the right shows the point when patient safety becomes compromised, after of which the occupancy continued to rise until peaking at almost 150% at 5 p.m. If crowding could be foreseen e.g. at 8 a.m. (dashed vertical line on the left), it would enable pre-emptive measures that would optimally result in EDOR levels shown in green line, and ultimately improved patient safety.
• Figure 2: Temporal distribution of crowded days among bedoccupying, medical and surgical patients in the sample. The two months of year 2020 are ommitted here for brevity. The figure demonstrates sporadic occurence of the crowding events that do not follow a deterministic weekday pattern, excluding the relative rarity of crowding during the weekends.
• Figure 3: Distribution of crowding events over different hours of the day among bedoccupying, medical and surgical patterns. Note That crowding is nonexistent between 8 a.m. to 11 a.m.
• Figure 4: Distribution of the crowding events during January 2018 as a function of hour of the day. Note the varying time of onset and duration of crowding as well as the relative rarity of crowding during the weekends.
• Figure 5: Area under the receiver operating characteristics curve (AUROC) and area under the precision-recall curve (AUPRC) at forecast origin 11 a.m. Guess level is provided in AUPRC plot for reference. 95% confidence intervals in parenthesis.