Short-Term Capacity Planning in an Intensive Care Unit: Web-Based Prediction Model and Pilot Study

• In ICU
• Tue, 14 Oct 2025

ICU Management & Practice, Volume 25 - Issue 4, 2025

ICU occupancy varies and cannot be scheduled. To tailor nurse staffing to variable ICU occupancy would potentially decrease both under- and over-staffing. We developed a simple web-based interface for short-term prediction of occupancy and, in a pilot study, compared predicted and observed occupancy.

Introduction

The capacity of intensive care unit (ICU) beds is limited. The main limiting factor for ICU capacity in Norwegian healthcare is often not funding, rooms or physicians, but the availability of intensive care nurses. A recent Norwegian governmental publication emphasised that the lack of healthcare workers will be the major future issue for Norwegian healthcare (Official Norwegian Report 2023:4). In Norway, the Regional healthcare authorities’ directives dictate that intubated ICU patients should have a nurse continuously at the bedside; thus, minimal capacity often equals the number of nurses. Nurses work every third weekend, which results in the number of scheduled nurses at each shift (day, evening, night) during the weekend being lower than during weekdays. In the current Norwegian labour market, the availability of intensive care nurses is limited. Therefore, to increase the workload with shifts every other weekend is not an option, as this will result in many nurses leaving the ICU workforce.

ICU patients are acutely ill, and it is not possible to plan a hospital stay or to postpone admittance. This results in a large variability of occupancy, and, ideally, the number of nurses should be tailored to the current need. However, it is not possible to have nurses' work schedules change on a day-to-day basis. Thus, the units are usually staffed in relation to the unit's defined number of staffed beds. For unexpected increases in the number of ICU patients, units have beds not staffed during normal circumstances and call in additional nurses. On the contrary, in periods with low occupancy, the unit may be overstaffed. Nurse overtime is expensive and also a stressor for the nurse workforce, which introduces a risk of increasing an already high personnel turnover. It would be beneficial to predict the expected occupancy during weekends in order to secure adequate staff and, at the same time, not use unneeded personnel resources.

Others have developed models for the prediction of ICU occupancy (Barado et al. 2012; Farcomeni et al. 2021; Ruyssinck et al. 2016). Many of these models are for long-term planning or for special circumstances such as pandemics. Our goal was to contribute to this work by developing a simple model for short-term occupancy planning easily applicable in day-to-day clinical practice and which can be introduced free of cost in ICU departments of different categories and settings. We developed and tested an algorithm for short-term prediction of ICU weekend occupancy based upon expected length of stay (LOS) for patients in the ICU and the expected number of new admissions during the weekend.

Methods

Part One: Development of Instrument

To predict the ICU occupancy during weekends at one day's notice (Thursday), three factors were included.

• For all patients in the unit on Thursdays, a physician predicted for each patient the expected additional ICU LOS (zero, one, two, three or four days or more). This will give an estimate of the expected burden related to patients currently treated in the unit. The prediction of expected LOS was both based on the patients' clinical condition and on organisational issues.
• Expected new admissions were calculated based on the mean number of daily admissions in the last year. New ICU patients' expected LOS was based upon data from last year, which specified the number of patients with a LOS of one day versus two days or more.  
• Some patients need more than one nurse due to high complexity. This can also be included in the model.

The project developed a web-based interface where physicians assessed the expected LOS for patients in the unit each Thursday. The model calculates and gives information about expected occupancy. The model was developed and implemented as a web-based interface, which can be customised for each particular unit.

Part Two: Validation of the Instrument

In a pilot period, the model was tested for its ability to give a prediction of weekend ICU occupancy. The test was performed in a mixed-case 10-bed ICU located in a 900-bed tertiary university hospital. The unit provided care for all categories of ICU patients except children, burn patients and cardio-thoracic postoperative care. The unit is a closed ICU where care of the patients is delivered by ICU physicians. The hospital has several high dependency units (HDU). Therefore, most patients in the ICU are dependent on advanced interventions such as intubation or renal replacement therapy.

In the pilot study, on each Thursday, an ICU consultant assessed the expected LOS for each patient in the ICU. The physician was instructed to consider all elements in this assessment, including expected clinical improvement, expected death and organisational issues related to patient discharge. Examples of the latter are the availability of air ambulance transfer to a local hospital; also, some HDUs in the hospital are closed during weekends. This assessment was the basis for expected occupancy during the weekend for patients who were in the unit on Thursdays. For occupancy of new patients, expected numbers were based on data from 2022, which showed that two patients would be admitted daily, of whom 58% would have a LOS of two days or more. Combined, these numbers gave a predicted occupancy on the weekend. During the weekends, the actual occupancy was registered daily. Both prediction and registration of actual occupancy were performed related to the number of patients in the unit at 12:00 AM.  

Statistics

Descriptive data were given for actual and predicted occupancy for all days and specifically for Friday, Saturday and Sunday. The assessment of agreement between predicted and actual occupancy was examined by several approaches. First, agreement at the group level between actual occupancy and predicted occupancy was addressed by the Wilcoxon Signed-Rank test. Second, difference scores for each day (difference = observed minus predicted occupancy) were calculated. A difference score within 1 was classified as good agreement, difference scores ≤ -2 overestimation (fewer patients than predicted) and difference scores ≥2 underestimation (more patients than predicted). Third, the strength of agreement between the assessments was reported using mean absolute error (MAE) and linear regression. This was a pilot study; therefore, no formal sample size calculation was performed. All analyses were done with the Statistical Package for the Social Sciences (SPSS) version 29.0.0.0.

Ethics

The project is a quality improvement project and therefore not within the scope of the regional ethical committee for health research and ethics. The study obtains no clinical data, and the researchers do not know the patient's identity. The project had no influence on the care provided to the patients. 

Results

Part One: Development of Instrument

The web-based instrument was developed in Anvil with integrated Python code (Anvil Works 2024 ). The computer code can be obtained from the authors. The interface for the user is available at https://cnnsnq5fy7zrz3ip.anvil.app/3XUJCO74GCSGDFJK6MC4IL7H (Figure 1). In this interface, the user can add each specific unit's beds. The user can also name each bed, adhering to local specifications. For each unit,  the expected number of daily new admitted patients and their expected LOS based upon organisational data are entered. The specified bed names and the numbers for expected new patients are included in future registrations until changed. In the weekly assessment, the physician checks off rooms that are occupied on Thursday and checks off for each day every patient is believed to stay in the unit. The interface also includes an option to enter the expected complexity of patients; for instance, in certain units, some patients only need 0.5 nurse, and some patients need more than one nurse. After this registration is done, the user clicks “Calculate”. The instrument then automatically produces a PDF file which, in addition to the entered data, specifies the calculated expected number of patients on Friday, Saturday and Sunday.   

The users completing the instrument were given a short verbal instruction, and the link was added to their browser bookmarks. Generally, users were able to complete the instrument in about 2-3 minutes. 

Part Two: Validation of the Instrument

In the pilot study, 14 weekends with a total of 42 predicted days were tested. The occupancies on Thursdays were a median of 8 patients (mean 7.4), ranging from 4 to 11. For the 42 days which were predicted, the mean predicted daily occupancy was 8.7 (SD 1.8) patients compared with the observed occupancy of 8.0 (SD 1.5) patients. The predicted number of patients was significantly different from the observed number of patients (p=0.02). The difference between predicted and observed number of patients for each weekday was not statistically significant (predicted vs observed Friday, Saturday and Sunday; p=0.19, p=0.15 and p=0.15, respectively).

The difference scores (observed minus predicted) ranged from minus 5 to plus 3. Twenty-one days showed good agreement, 15 days overestimation (fewer patients than predicted) and 6 days underestimation (more patients than predicted). The prediction for Fridays was more precise (overestimation 3, good agreement 11, underestimation 0) compared with corresponding numbers on Saturdays (4-7-2) and Sundays (7-3-4). The distribution of difference scores for all days and for each weekday is given in Table 1.  

The mean absolute error (MAE) difference between predicted and observed number of beds was 1,6 (95% CI: 1.2-1.9). Linear regression between observed and predicted number of beds demonstrated no statistically significant association (P=0.13). 

Discussion

This pilot study showed that short-term prediction of weekend occupancy based upon expected stay for patients in the unit on Thursdays and expected number of new admissions during the weekend, in general, overestimated the need for beds during the weekend. Based upon the predictions, the unit would have been understaffed for only a few days.

This study based expected occupancy on the consultant ICU physicians' consideration of expected LOS for each patient within the unit on Thursday. The assessment is based on a combination of factors related to the patient and the organisation. Relevant patient factors are severity of disease, organ failure and functional status. For instance, it is unrealistic that a patient in prone position and high settings for ventilator pressures will be weaned within the next few days. However, increased severity of disease does not automatically imply longer LOS. The patient may, due to severe disease, be expected to die within a short time or may, due to complexity, be planned for transfer to another ICU or hospital. Organisational factors are often related to local hospital capacities; other step-down units may have no free beds or may be closed during the weekends. Which patients are transferred to other units is also different between units; for instance, a unit in a small-volume hospital will often transfer patients to tertiary hospitals. Overall, this suggests a large number of potential predictors with multiple potential for interactions.

Gonzáles-Nóuva et al. (2023) applied artificial intelligence for ICU occupancy estimation. They included data from sensors, medical history and clinical-chemistry results exported from the electronic medical record in the gradient boosting technique for learning. Using a large number of observations, they completed a model that could predict LOS with an MAE of 2.5 days in a unit with a LOS average of 4.3 days. This study estimated LOS, not the number of beds on specified days, and is therefore not directly comparable with our observations. The relatively high MAE suggests that the artificial intelligence model did not perform much better than physicians’ assessments. This may be because experienced ICU physicians, consciously or non-consciously, in a heuristic thought process, include the relevant factors when determining expected patient trajectories. Ryussinck et al. (2016), who also used artificial intelligence, found that respiratory organ failure and current LOS are more important for future trajectory than coagulation organ failure. This would be obvious to an experienced clinician.

Perhaps occupancy prediction could be made simpler. Jin et al. (2021) developed a simple formula based on historic 12-month activity data and linear regression. Their department is similar to our unit, an emergency ICU caring for most categories of patients. In their unit, with 12 ICU patients and 20 HDU patients, they calculated the upcoming need for nurses the next day as N = (0.45 x number of nurses needed 24 hours prior) +11. With this simple formula, they had 14% of days with less than minimum staffing and 14 % with more than three extra staff (Jin et al. 2021). These data are comparable to our results of 6 of 42 days (14%) with more than one patient in excess of predicted and 7 of 42 days (17%) with three or more nurses in excess.

From the results cited above, it seems that a complex model using artificial intelligence and our simple model using physicians' assessment, which considered the number of current patients, did not differ much in performance. However, these studies were from different populations, and thus, it is conceivable that a different case mix may have influenced the results.

What are the practical potential implications of the prediction of ICU occupancy? As shown in our study and in other studies, it is not possible to do an accurate estimation of future occupancy. Some uncertainty will exist because of the obvious lack of a schedule for incidents such as cardiac arrests, onset of severe sepsis and traffic accidents. Still, our data suggest that the unit could, for a number of days, expect a lower occupancy than normal and thus have a lower than standard staffing. This can be done by not replacing personnel on sick leave or voluntarily moving personnel to other working days. Another model, as proposed by Jin et al. (2021), is to reduce the standard staffing on weekends and establish an on-call system for nurses, which titrates the needed workforce. Even with financial incentives for nurses to be on-call, this organisation lowered the total cost for the unit. 

Our data shows that the expected occupancy was more often overestimated than underestimated. Thus, the risk of being severely understaffed was low. Of the 42 days, 5 days had one more patient than expected which is usually easily managed, 5 days with two more patients than expected which can be managed on a short term-basis using, for instance, staff otherwise occupied with postoperative care, and only one day with three more patients than expected which may require a more urgent need for calling in nurses from their home. On the other hand, the overestimation could lead to being overstaffed unnecessarily on some days. Our data suggest that the department with this model would be staffed in relevant excess for 7 of the 42 days. This may be acceptable in an emergency organisation. Moreover, the nurses would be in excess in relation to the ICU's needs. In many hospitals, including ours, nurses in excess in one unit may contribute to work in other units, such as postoperative care units or paediatric ICUs.

Hospital managers are usually subject to financial budget strain. Thus, there may be an administrative pressure to reduce scheduled ICU staffing. This may be a short-sighted strategy as it may increase workload for nurses, increase the number of unexpected shifts on weekends and due to lack of beds, result in time-consuming administration of patient flow. All these factors are recognised to increase burnout among ICU personnel and staff turnover (Poncet et al. 2007). Recruitment of ICU personnel is time-consuming and costly, as new personnel usually need a training period before they are able to work independently of support from others. Therefore, to minimise staff can paradoxically add costs. To have a model where staff is reduced not by a lower scheduled standard staffing but at a time when low expected occupancy is safe, reduces the burden for employees working on weekends and also reduces costs. 

It is important to underline that each model must be developed according to specific ICU department characteristics. Some ICUs admit only emergency patients, which cannot be planned, while some units admit patients from elective surgeries, which can be planned, and, if needed, rescheduled. Some units have a high patient turnover, while other units, such as neurosurgical ICUs, have generally longer LOS. Some hospitals have an HDU acting as a step-down unit, which improves patient outflow from the ICU. Finally, medical interventions and LOS may vary between units. There are also a number of readmissions to the ICU. This number varies based on the medical capacity in the wards. 

Our model was a simple one and is therefore to be considered perhaps more as a proof of concept. We did not include the expected complexity of patients; some patients were in need of more or fewer than one nurse. Scoring systems, such as the Nursing Activities Scoring (NAS) system, can quantify the need for nurses but will involve a more time-consuming scoring process (Stuedahl et al. 2015). However, to differentiate between patients is more relevant in units with a mixed HDU and ICU patient population and was therefore less relevant in our unit. To accommodate other units, the web instrument is also designed to register the number of nurses needed for each patient included. Another factor not included is the within-day variability. There will usually be highs and lows for occupancy during each day. To what degree this variability exists depends on the characteristics of the unit. In units with high patient turnover, such as emergency departments, diurnal variability may be high; in more long-term facilities, very low. 

We recognise that this study has some limitations. First, it is a pilot study with a limited number of observations. Second, it is a one-centre study, which limits its external validity. The study has also not reported detailed clinical characteristics, making comparisons to other settings difficult. Finally, this instrument is tested during normal working conditions. Special circumstances, such as the COVID-19 pandemic, need more complex forecasting, which includes data from the progression of disease in the general population (Redondo et al. 2023).

In conclusion, we demonstrate that a simple instrument to short-term predict ICU occupancy performed similarly to more complex models. The prediction errors were generally skewed towards overestimation and did not result in a critical lack of nurses. This instrument can aid ICU managers to more precisely tailor weekend nurse staffing to occupancy. 

Conflict of interest

None

References:

Anvil Works. Available from: https://anvil.works. 

Barado J, Guergué JM, Esparza L, et al. A mathematical model for simulating daily bed occupancy in an intensive care unit. Crit Care Med. 2012;40:1098-104.

Farcomeni A, Maruotti A, Divino F, et al. An ensemble approach to short-term forecast of COVID-19 intensive care occupancy in Italian regions. Biom J. 2021;63:503-13.

González-Nóvoa JA, Busto L, Campanioni S, et al. Two-step approach for occupancy estimation in intensive care units based on Bayesian optimization techniques. Sensors (Basel). 2023;23:1162.

Jin Z, Jovaisa T, Thomas B, Phull M. Intensive care unit staffing during the periods of fluctuating bed occupancy: an alternative dynamic model. Intensive Crit Care Nurs. 2021;66:103063.

Official Norwegian Report (NOU) 2023:4. Tid for handling — Personellet i en bærekraftig helse- og omsorgstjeneste. Available from: https://www.regjeringen.no/no/dokumenter/nou-2023-4/id2961552/?ch=1. 

Poncet MC, Toullic P, Papazian L, et al. Burnout syndrome in critical care nursing staff. Am J Respir Crit Care Med. 2007;175:698-704.