Predicting Need for Hospital Beds to Reduce Emergency Department Boarding

Abstract

Introduction

Accurately determining the supply and demand of hospital beds for new admissions can help prevent adverse patient outcomes. Quantitative analysis of modern electronic medical record data can help predict supply and demand for unoccupied staffed hospital beds (SEDs) and aid in eliminating human approximations, standardizing daily work through concrete and objective data. The purpose of this study was to reduce variability and human error in predicting the number of SEDs needed.

Methods

In this study,the authors analyzed bed calculator data from a medium-sized, suburban medical center to evaluate the efficacy of a unique bed calculator prediction tool to determine the need for SEDs. The calculator aggregates multiple key reference factors available through the bed calculator system into a cohesive linear regression model.

Results

Compared with human estimation, the authors found that the bed calculator is able to predict the number of SEDs needed more effectively. That being said, there was no significant difference in the average boarding times pre- and postintervention, indicating that the bed calculator did not result in decreased boarding times for patients.

Discussion

These findings establish the efficacy of the bed calculator and its ability to align bed supply and demand. Because patient boarding times depend on the system’s patient flow management, future studies should focus on how to improve various streams of communication and coordination.

Introduction

One of the major indicators of better hospital operational management is accurately determining the supply and demand of hospital beds for new admissions. Accurately measuring and predicting the patient census is particularly important in a hospital, as inaccuracies can result in significant adverse outcomes. For example, Singer et al found that longer emergency department (ED) boarding times are associated with negative patient-oriented outcomes, such as longer hospital stays and higher inpatient mortality rates. ¹ This suggests that efforts to reduce boarding times may improve outcomes for patients who are admitted from the ED to the hospital. Although hospitals may then prefer to overstaff nurses to ensure a high quality of care, overstaffing can result in unfavorable outcomes, such as unnecessary costs and stagnant capital. ² Given the rising costs of inpatient care and constrained resources of hospital systems, overstaffing cannot and should not be common practice. Studies in the US and other countries have indicated that patient experience in the ED is a growing area of focus for leaders throughout health care. Decreasing wait time does more than reduce costs. Wait time is a major driver of the patient experience. Prolonged wait times at the ED are not only associated with increased morbidity and mortality, but also with decreased satisfaction. ^3–6 Patients who are satisfied with the care they receive are then more likely to comply with their treatment and access resources within other departments of the hospital. ⁷ A major contributor to the supply–demand mismatch and resultant ED crowding is the length of ED boarding, where longer boarding times are associated with increased hospital mortality and longer lengths of stay. ¹ Therefore, admission boarding times present a significant threat to patient health outcomes that needs to be addressed.

Efficient patient hospital admission requires a high degree of coordination and communication among different sectors in the hospital, and health systems across the country have adapted strategies to reduce ED boarding times. In general, patient census is the most important factor for hospital bed capacity planning. ⁸ Chang et al compared hospitals in the top and bottom 5% of the 2012 Centers for Medicare & Medicaid Services adjusted length of stay and boarding times for admitted patients. ⁹ It was found that high-performing hospitals used data-driven strategies to specifically change behavior and predict hospital flow to match resources with hospital needs accurately, rather than relying on variable staff feedback. ⁹ Modern electronic medical record systems are troves of information that can shed light on hospital operations. ¹⁰ Quantitative systems using bed calculator data to predict the supply and demand of unoccupied staffed hospital beds (SEDs) can aid in eliminating human approximations and systemize daily work through concrete and objective data. (An astute reader may question why SEDs exist? This situation can occur for various reasons. For example, a patient may discharge, the nurse is still present, and the room is being cleaned. That bed will soon be ready for a patient currently boarding in the ED. It is important for any supply–demand model to account for this type of situation, and not call for unnecessary additional staffing.)

In this study, the authors analyze bed calculator data from a medium-sized, suburban, medical center to evaluate the efficacy of a unique bed calculator prediction tool to determine the need for staffed beds. Prior to the study, the need for staffed beds was largely determined by the individual house supervisor’s assessment of hospital needs. The house supervisor’s duty is to assign unoccupied staffed hospital beds to the admitted patients. This individualized determination introduces variability into the system and can result in inaccurate assessments and understaffed or overstaffed beds. The system will aggregate multiple key reference factors into a cohesive linear regression model, including SEDs, number of boarders in the ED, pending hospital-based specialist (hospitalist) consults, total ED census, pending transfers, adult surgical admits, hospital discharge orders, complex hospital discharges, and additional relevant factors available through the bed calculator system. The application of these factors will produce a more standardized prediction model for bed capacity planning.

Methods

Statistical model

Staffing dynamics vary from hospital to hospital. Our hospital has 3 changes of shift per 24 hours at 7:00 am, 3:00 pm, and 11:00 pm. In the authors’ hospital, staffing is finalized 4 hours before each change of shift. Thus, our model’s prediction window is 4 hours. The authors collected 2 months of preliminary data at every change of shift and 4 hours before every change of shift to document the house supervisor’s accuracy and test the concept that a statistical model could predict the number of beds needed 4 hours in the future.

Data collected included every type of movement into or out of the authors’ hospital: SEDs (ie, beds ready for a patient or being cleaned and ready soon with resources available), boarders in ED (ie, patients with admit orders), pending hospital-based specialist consults, total ED census, pending transfers, adult surgical admits, hospital discharge orders, complex hospital discharges, and the house supervisor’s prediction of number of beds needing to be staffed in 4 hours.

Our hospital’s goal is for admitted patients to transfer to a hospital bed within 60 minutes. The house supervisors are adept at filling any SEDs. The challenge that the bed calculator addresses is predicting the change in staffing needed for the next shift to have an appropriate number of SEDs. The authors defined the bed deficit = boarders in ED − SEDs. The goal of the prediction and subsequent staffing is to match supply and demand perfectly, with a bed deficit of zero during the next shift. Because the authors wanted a model that was easily explained to increase staff acceptance, linear regression was used to predict the bed deficit 4 hours into the future. Pending transfers, surgical admits, and complex hospital discharges were found not to be statistically significant predictors, reflecting their low volumes. The linear regression analysis in Excel calculated that bed deficit is nearest to 0 when

Predicted bed deficit = −0.76*SED + 0.99*boarders in ED + 0.49*pending hospital-based specialist (hospitalist) consults − 0.32*hospital discharge orders −4.44

This equation can be explained to staff in simple terms as follows: Probably 76% percent of unoccupied staffed beds will be filled in the next 4 hours, thereby reducing the bed deficit. Ninety-nine percent of boarders in the ED will fill hospital beds in the next 4 hours. Forty-nine percent of pending hospital-based specialist consults will fill hospital beds in the next 4 hours. Thirty-two percent of hospital discharges will result in an unoccupied bed in the next 4 hours. The constant 4.44 is a correction factor the linear regression calculated to keep the equation unbiased with the predicted bed deficit as close to 0 as possible.

The authors could then compare the house supervisor’s prediction with the model’s prediction.(Table 1)

Table 1:

Accuracy of predictions

Prediction vs observed number of beds 4 h later	Bed calculator	House supervisors
12 or more over	0.034483	0.034483
8 over	0.043103	0.034483
4 over	0.12931	0.060345
Correct	0.551724	0.215517
4 under	0.146552	0.155172
8 under	0.060345	0.137931
12 or more under	0.034483	0.362069
Overall	83% of predictions are correct within 4 beds. Mean prediction is correct	44% of predictions are correct within 4 beds. Mean prediction is 8 under

The house supervisor’s prediction had a mean of 8 beds less than needed, which results in patients boarding unnecessarily in the ED each shift. Thus, replacing the house supervisor’s estimations with the proposed bed calculator predication could improve resource matching. It was also noted that the accuracy of the bed calculator was nearly double that of the house supervisor, resulting in less than 8% of shifts overstaffed. The authors applied to the institutional research board for and received a not human subject’s research determination prior to using the bed calculator’s predictions.

Data set Description

The data collected were adult-only medical–surgical census numbers counted and collected from the Clarity (HealthConnect backup). Patients in the intensive care unit were excluded because they have more complex dynamics (upgrades and downgrades), are staffed at a different nursing ratio, and are a relatively small population (20 vs 146 licensed beds). The bed calculator was first introduced in May 2020, so the control arm consists of 6 months of data from October 2019 to March 2020. The experimental arm is a 6-month period from December 2020 to May 2021.

Bed Calculator Implementation

The bed calculator is an aggregate of multiple, real-time reference factors and calculates the predicted bed deficit. Knowing the change in number of staffed beds needed 4 hours in advance of the change of shift assists the house supervisor to match supply and demand. The house supervisor records in an Excel spreadsheet the number of SEDs, boarders in the ED, pending hospital-based specialist consults, and hospital discharge orders. The equation embedded in the spreadsheet calculated and displayed the number of beds predicted to be needed 4 hours in advance of the change of shift. This information was shared with the staffing office.

Goal calculation and analysis

The outcome variable used to measure the impact of using the bed calculator is called the “Goal” variable. The Goal is a binary indicator variable with 1 being the goal has been met and 0 being that it has not. In this study, the Goal is met if the SEDs or boarders in the ED values are less than 2 (Figure 1). In our hospital nurse-to-patient staffing ratios are 1:2, 1:4, or 1:5, so the goal less than 2 corresponds to staffing being accurate to a single nurse or better. This measured whether or not the bed calculator is successful in accurately predicting the number of beds needed 4 hours in advance. A higher score indicates greater alignment between SEDs and boarder in the ED values. One hour after each change of shift (8 am, 4 pm, and midnight), the Goal was calculated and the pre- and postintervention averages were compared. Figure 1 describes the overall data analysis process.

Figure 1:

Goal variable analysis process. The goal is calculated each day at each time point (8 am, 4 pm, and midnight) and the average Goal values are summarized for control and experimental group comparison periods. BC = bed calculator; BED = boarders in the emergency department; SED = unoccupied staffed hospital beds.

Correlation between experiment time and hospital COVID census

The first COVID-related patient death occurred in the hospital of study in March 2020. Since then, the COVID patient census has fluctuated significantly from month to month. In order to accommodate the COVID patients’ isolation needs, ordinary hospital beds were not always suitable for such patients. The authors did not anticipate the large numbers of isolation patients associated with the pandemic, when the bed calculator was created. In order to avoid the complications introduced by the COVID surge, the authors chose a time frame where the number of COVID patients were relatively stable (Figure 2). ED census also fluctuated during the pandemic, as shown in Figure 3.

Figure 2:

Daily average inpatient COVID census. Monthly average patient census from March 2020 to January 2022 was calculated and graphed.

Figure 3:

Monthly emergency department census. Total monthly patient census from March 2020 to January 2022 was graphed.

Results

Goal results

In order to determine the efficacy of the bed calculator on accurately predicting the number of SEDs needed 4 hours in advance, the daily Goal (boarders < 2 or beds < 2) was calculated at each time point (8 am, 4 pm, and midnight). The average Goal score pre- and postintervention for each time point was summarized and graphed (Figure 4). The goal score was significantly higher postintervention at the 8 am (M_Pre = 0.19, SD_Pre = 0.08; M_Post = 0.76, SD_Post = 0.08) and midnight (M_Pre = 0.06, SD = 0.06; M_Post = 0.29, SD_Post = 0.09) time points (p_8am = 0.00, p_midnight = 0.01; α = 0.05). These data suggest that, at 8 am and midnight, the bed calculator is able to improve the prediction of the number of SEDs needed significantly, 4 hours prior.

Figure 4:

Bed calculator accurately predicts number of unoccupied staffed hospital beds needed for 8 am and midnight time points. Average goal scores pre- and postintervention were calculated at each time point (8 am, 4 pm, and midnight) and graphed. Error bars represent the standard deviation of each time period and stars above the postintervention bars indicate statistically significant differences from the preintervention counterparts. MN = midnight.

Patient boarding time results

In order to determine the impact of the bed calculator on the patient boarding time, the boarding time pre– and post–bed calculator introduction (collected from bed calculator data) were summarized and graphed (Figure 5). There was no significance in the average boarding time pre- and postintervention, indicating that the bed calculator did not result in decreased boarding times. This unexpected result may be due to other factors influencing ED and hospital flow during the observation period, such as surges in the COVID-19 pandemic. For example, the hospital was frequently at maximum capacity. In that situation, the bed calculator predicting more need was not actionable.

Figure 5:

No significant impact on patient boarding times after introduction of bed calculator. Monthly average patient boarding times before (preintervention) and after (postintervention) the introduction of the bed calculator were calculated and graphed. Error bars represent the standard deviation of each time period and no statistical significance was found via t-test.

Discussion

In this study, the authors examined the ability of the bed calculator, a novel bed prediction tool, to predict the number of staffed hospital beds needed for admitted patients from the ED, 4 hours in advance. By comparing bed calculator data pre– and post–bed calculator implementation, it was found that the bed calculator is able to predict the number of beds needed more effectively, compared to relying on house supervisor estimates.

These findings somewhat align with those of Chang et al, ⁹ who performed a comparison of hospitals in the top and bottom 5% of the 2012 Centers for Medicare & Medicaid Services adjusted ED length of stay and boarding times for admitted patients. The authors found that 1 of the most important factors in improving ED boarding times is a data-driven strategy that can more accurately match resources with hospital needs rather than relying on staff feedback, which can be variable depending on the person. ⁹ This is clearly demonstrated at the 8 am and midnight time points during the postintervention period, which yielded significantly higher Goal averages compared to the preintervention period. The lack of statistical significance at the 4 pm time point could potentially be explained by the relatively higher volume of patients at that time, often occupying every bed, and variations in placement priorities at different time points. Although the data-driven bed calculator approach has resulted in greater alignment of bed supply and demand, it has not resulted in decreased ED boarding times for admitted patients.

Because patient boarding times depend on the system’s patient flow management, many external factors during the process could influence the overall outcome. One reason the bed calculator failed to reduce the overall boarding times could be related to a bottleneck elsewhere, although the hospital at maximum capacity was an obvious limitation. If the hospital census returns to pre-COVID levels, a future study could characterize the system’s patient flow management and identify opportunities for improvement. For example, could our discharge process, which averages more than 4 hours, be expedited? Such an approach was taken in Alhaider et al, who used the Distributed Situation Awareness framework to identify deficiencies and found that the framework is effective in analyzing communication and coordination processes. ¹¹ It is possible that system reform to better patient flow management is needed to see improved patient boarding time in the future.

Another reason there failed to be a marked difference in patient boarding times could be the temporal resolution of the authors’ data collection. The number of boarders was recorded every 4 hours. Access to more granular, hourly variations in boarding times could paint a richer picture of bed calculator impact. Because of lack of access to hourly data points, the authors were unable to determine the hourly patient boarding time and had to average monthly data, which provided a partial, yet incomplete picture of the impact of the bed calculator. Future improvements on the bed calculator should focus on both identifying ways to improve the patient flow management and developing a more comprehensive data collection methodology. This study was based on a single midsize suburban hospital study. Generalization would require studies at other hospitals including different possible influencing factors, such as the size of hospital, patient demographics and population, as well as geographic settings might shed more light on how to modify the bed calculator equation further. Furthermore, this study was designed prior to the COVID-19 pandemic, and outcomes could be compared once the pandemic has passed.