Evaluating the impact of emergency department crowding on disposition patterns and outcomes of discharged patients

We conducted a retrospective analysis of coded data from electronic health records (EHRs) of patients evaluated at an adult ED (where the majority of patients are aged 18 years and older) over a 20-month period between the years 2012 and 2014.

The study site is an 86-bed ED located in a tertiary care teaching hospital in an urban setting. The annual ED patient volume was approximately 85,000 during the study period. The ED is staffed around the clock with board-certified attending emergency physicians and supervised physician assistants and trainee resident physicians.

This study was approved by the Institutional Review Board at our institution.

We used data from all patients who received care in the ED during the study period.

The primary outcome variable was patient disposition; whether or not the patient was admitted (including either to observation or inpatient settings) or discharged. As secondary outcomes, we also examined, for those who were not admitted, the indicator of return visit to the ED within 2 weeks of discharge and, among those that did return within 2 weeks, we examined whether they were admitted during their return visit.

The primary exposure was whether or not the patient’s disposition was decided during a high occupancy hour. Because occupancy can vary substantially throughout a given day, the determination of high occupancy was based on the total occupancy (in terms of patient-per-bed) during the hour of disposition decision. To convert this occupancy to a proportion, we used the total number of available beds in our ED in each hour as the denominator. At our facility, the number of available beds fluctuates in a predictable manner during the course of the day (i.e., some beds are closed on a pre-planned basis at the same time each day), hence the occupancy denominator is variable during a 24-h cycle. We used the total number of available beds in our ED in each hour as the denominator. ED occupancy in each hour was compared to that same hour on all other days in the data set, and all hours in the top decile of that list were designated high occupancy hours, producing a binary indicator of high occupancy. We compared each hour to only the same hour across all days. For example, if an individual’s disposition decision occurred at 7:55 pm, then the ED occupancy between 7:00–8:00 pm on that day was compared to the occupancy during the same timeframe on all other days in the dataset to determine if it was in the top decile. In this way, each hour of a given day is marked as high occupancy independently of the other hours of the day. We took this approach, instead of comparing each hour to all other hours in a given day, because we noted that when we compared to all other hours, some AM hours were never marked as crowded, since they were always at a lower occupancy than other more typically crowded hours of the day. Further, comparing each hour to the same hour across all days in the dataset will allow the impact from factors such as time during a shift (e.g., beginning, middle, and end of a shift) and the time before and around sign-out to be held constant—as such factors may impact disposition decision-making independent of ED occupancy levels or presence or absence of boarders.

All patients who were placed in a bed in the ED are captured as part of the total occupancy (regardless of whether they left with or without being seen). However, only patients who had a discharge or admission diagnosis were included in the analysis of disposition decisions and assessment of the study outcomes.

ED occupancy fluctuates according to the time of day, day of the week, and the time of year. We chose to measure the ED occupancy every hour to capture this variability. The more granular the measurements of crowding, the more information is preserved to study the effects of crowding on the outcomes of interest. In particular, occupancy that is measured at a daily level may mask much of the variation in occupancy that occurs within a 24-h period [50]. For example, a patient whose disposition decision is made at 7:00 am when the ED is not crowded is not affected by crowding that occurs later in the day. However, a model that measures occupancy in 1-hour increments would allow assessing whether patients’ disposition decisions occurred during a crowded period. Furthermore, from a pragmatic standpoint, a 1-hour interval likely better reflects the reality of working in the ED. The physician is unlikely to be aware of minute-to-minute fluctuations in the ED occupancy, or 24-h fluctuations. As a final rationale, a supplemental analysis of data from our ED showed that approximately 55.2% of the total variation in occupancy was due to hour of day, 22.4% was due to daily variation, and 22.4% was unexplained variation (Appendix).

The other exposure of interest in this analysis was the presence of ED boarders. ED boarding is defined as holding an admitted patient in the ED until an inpatient bed becomes available [51]. To identify hours with high boarder patient counts, we first calculated boarding time for each encounter as the difference between ED disposition and inpatient (IP) floor arrival time. Because these two times are never exactly equal (i.e., every encounter has some non-zero boarding time), we only counted a patient as a boarder if his boarding time was above the average of 4.29 h for our sample. Because this still identified > 90% of all hours as having boarders, we further only classified an hour as having boarders if that hour had an above average number of boarders (8.69). Using this method, we increased the likelihood of identifying high occupancy ED hours that correspond with low IP capacity.

We began with descriptive analyses of the sample based on patient demographics, severity of illness, payer source, day of week and arrival season, number of ED to ED transfer denials on arrival day, percentage with boarders at disposition decision hour, and percentage with disposition decided during a high occupancy hour. This analysis was then stratified by those who were (vs. were not) admitted, and two-sample comparisons were made to conduct an unadjusted evaluation of variables associated with admission.

Using a logistic regression model predicting admission, we obtained estimates of the effect of high occupancy on admission disposition, adjusted for age, race, gender, severity of illness (using facility level billing [52] and the Charlson index [53, 54]), payer source, day of week, season, presence of boarders, and number of ED-ED transfer denials. The Charlson-Deyo index is a commonly used co-morbidity index, assessed from each patient’s discharge diagnosis ICD-9 codes, that is predictive of hospital mortality [53, 54]. Facility billing level is a CMS outpatient payment coding system analogous to inpatient Diagnosis Related Groups (DRGs). It is used to reflect the volume and intensity of resources utilized by the facility to provide patient care [52]. A measure of high boarder patient counts was also included in all models. In a second analysis, we explored the role of high boarder patient counts as a possible modifier of the relationship between high occupancy and admission by presenting models stratified by presence of high boarder patient counts. We also fit logistic regression models to examine how high occupancy at disposition decision hour correlates with 14-day return ED visits for those who were discharged and, among those that did return, whether they were admitted during the return ED visit.