Prolonged mechanical ventilation in critically ill patients: epidemiology, outcomes and modelling the potential cost consequences of establishing a regional weaning unit

We performed a retrospective cohort study using prospectively collected data available in the Scottish Intensive Care Society Audit Group (SICSAG) database [10]. The Local Research Ethics Committee (Lothian Research Ethics Committee 3) waived the need for formal ethical review. Patient confidentiality was ensured as the dataset was fully anonymised.

Our study was based in Lothian Health Board, which serves a population of 900,000 in South East Scotland. The region is served by three adult hospitals managed as a single organisation providing general ICU services as well as some national and specialist regional services: neuroscience (including traumatic brain injury), solid organ transplantation (excluding heart/lung), trauma, acute liver failure, and complex vascular surgery. Each hospital has a "closed" adult general mixed medical/surgical ICU with specialist intensive care staff. The number of funded ICU beds in the region varied from 23 to 26 over the study period [see Additional file 1, Table S1]. In total, the three ICUs typically manage about 1,100 MV patients per year. Referral patterns for South East Scotland mean that ICU patients comprise almost all patients requiring critical care in the geographical region, as well as some patients from outside the region for whom national or wider regional specialist services are provided. Cardiac surgery and coronary care services are managed in separate units and are not included in this analysis. Paediatric services are also not considered.

Every patient admission episode to an adult general ICU in Scotland generates a record that is stored in the SICSAG database [10]. A national validation study showed that the data quality was of a high standard when compared with clinical case notes, with only a 6% level of disagreement [11]. All data are entered prospectively by clinical staff and include demographics, Acute Physiology and Chronic Health Evaluation (APACHE) II scoring, and diagnostic data. Organ support data (e.g. invasive ventilation, renal replacement therapy (RRT), vasoactive therapy) and presence of a tracheostomy are entered on a daily basis during an admission episode. RRT in the three units is usually continuous haemofiltration, although intermittent haemodialysis is occasionally provided for more stable patients. Non-invasive ventilation is available in all units, but is not recorded in the SICSAG database. Survival status at ICU and hospital discharge is also recorded.

All patient records were extracted from the SICSAG database from the three ICUs over a five-year period, from 1 January 2002 to 31 December 2006. The following two groups were excluded prior to analysis: patients under 16 years old and patients transferred between ICUs with incomplete data for at least one admission episode. These incomplete admission episodes resulted from patient transfers to/from ICUs that did not contribute to the Lothian regional dataset.

Comorbidities were obtained from APACHE II chronic health evaluation categories. Admission diagnosis was derived from APACHE III diagnostic categories, with new groupings created by manual review by one of the authors (NL) for pneumonia, trauma, and sepsis [see Additional file 1, Table S2]. Outcome variables were recorded in the SICSAG dataset. No assumptions were made regarding missing data, and analyses were undertaken using a complete-case analysis.

Calculation of incidence requires an agreed numerator and denominator, and neither are universally agreed when describing the epidemiology of PMV. For numerator data, we used two definitions of PMV to provide sensitivity analyses given the potential limitations of retrospective data as detailed below.

The number of bed-days used by the PMV population was calculated as a proportion of total funded level three bed-days in the three Lothian ICUs over the study period. A level three bed is one in which a patient receives MV and/or multi-organ support. Annual estimates of number of funded ICU bed-days in Lothian were obtained from local service managers.

We limited modelling to data from the years 2005 and 2006 because consecutive daily organ support data were available for this subgroup. Four different hypothetical weaning units were modelled using different admission criteria (see Table 1). In all cases we used consecPMV as the numerator.

Criteria were chosen after discussion with intensive care clinicians locally and a review of the literature, acknowledging that weaning units were unlikely to admit patients still requiring cardiovascular support, and may not be able to provide RRT. In general, unit A represented the most stable patients requiring only ventilatory support who were likely to have a low risk of clinical deterioration. In contrast unit D represented patients requiring ongoing renal support in whom cardiovascular support was only recently discontinued, and who may have been at greater risk of clinical deterioration. Units B and C were modelled to include patients at intermediate levels of risk.

For modelling, the analyses were undertaken using syntax programming in the statistical software package SPSS (SPSS Inc., Chicago, IL, USA), and the results were crosschecked manually for unit A in a spreadsheet package (not shown). The date of eligibility for each PMV patient for each of the four hypothetical units was determined, excluding patients who never reached eligibility criteria. In this eligible patient group, no patient had more than one admission episode. For each hypothetical unit, baseline characteristics of eligible patients were described, together with ultimate hospital outcome. We used the remaining length of stay in ICU after eligibility from individual level patient data to calculate occupancy rates, varying capacity from one to eight beds for each of the four weaning units. We also estimated the likely refusal rates for admission as a result of inadequate capacity by assessing weaning unit bed availability on the day that a patient became eligible for admission to the weaning unit. If the unit was full on this day, this was counted as a refusal. Costs were calculated using a top-down approach which included physician and staff costs. Prices were converted from UK£ to euros (€) using the Organisation for Economic Co-operation and Development Purchasing Power Parity rate of €1 = UK£0.822292 [12]. An ICU bed-day was assumed to cost €1,690 per day (£1,390 published UK National Health Service (NHS) tariff 2009/10) [13], and a weaning unit bed around half this value [9]. The reduced cost of weaning unit beds is due to the expected lower staffing ratios required. A sensitivity analysis was undertaken varying the cost of a weaning unit bed from 50% to 100% of the cost of an ICU bed. The model assumed that staffing an unoccupied bed in the weaning unit would accrue the same cost as an occupied bed. In addition, eligible patients would only be transferred once a bed became available.

All analyses were undertaken using SPSS v14. Patient characteristics were presented as number and percentage, mean and standard deviation (SD), and/or median and interquartile range (IQR). Both mean and median values were reported for continuous variables with skewed distributions, which were important in reporting resource use, for example length of ICU stay. Characteristics were described for PMV and non-PMV groups, and compared with the following tests: t-test for normally distributed data, Mann-Whitney U test for non-normally distributed data, and chi-squared test for categorical variables. Trends were analysed using chi-squared test for trend for categorical variables. The association between PMV status and diagnostic category was assessed using relative risk, although only those with a P value less than 0.0004 were considered to be significant to correct for multiple comparisons. Confidence intervals (CI) for incidence rates were derived using the Poisson distribution [14]. A significance level of 5% was used for analyses and 95% CI were presented (unless stated).

Patients who required PMV (defined as countPMV) when compared with non-PMV patients for the years 2002 to 2006 were older, more likely to be non-surgical admissions, had higher overall severity of illness on admission, and worse oxygenation on day one, but had fewer recorded comorbidities (Table 2). Pneumonia, septic shock, and trauma were among the most common five reasons for admission in both groups. However, only five ICU admission diagnostic categories were associated with a significantly higher risk of PMV (P < 0.0004 for all comparisons): Guillain-Barré syndrome (relative risk (RR) = 12.2), pancreatitis (RR = 4.6), acute respiratory distress syndrome (RR = 4.0), pneumonia (RR = 3.4), and septic shock (RR = 2.1).

A comparison of outcomes for PMV and non-PMV groups is shown in Table 2. PMV patients had a non-significant increase in ICU mortality (absolute difference = 2.9%, 95% CI -1.9% to 8.3%, P = 0.23), which reached statistical significance at hospital discharge (absolute difference = 6.5%, 95% CI = 1.1% to 12.2%, P = 0.02). PMV patients had longer post-ICU hospital stays than non-PMV patients (P < 0.001). Patterns of discharge from the acute hospital were similar for the two groups (P = 0.06), with a high proportion of PMV patients discharged to their own residence from the acute hospital (89%) and small numbers being transferred to rehabilitation or long-term care facilities (11%).

The incidence of PMV calculated using different definitions of numerator and denominator is shown in Table 3. For countPMV (data 2002 to 2006) the incidence was 4.4 per 100 ICU admissions using all admissions as the denominator and 6.3 per 100 ventilated ICU admissions. For consecPMV (data 2005 to 2006) the incidence was 3.7 per 100 admissions, which compared with an incidence of 3.9 per 100 admissions when countPMV was calculated for the same period.

Over the entire study period (2002 to 2006) there were, on average, 24.4 funded level three ICU beds per year in the region (Table 1), and there was a mean of 70 PMV cases annually. Mean ICU length of stay per PMV case was 37.2 days, equating to 29.1% of all funded level three ICU bed days over the study period.

Out of 126 patients requiring PMV during the period 2005 to 2006, between 80% (unit A) and 93% (unit D) were eligible for transfer to the hypothetical weaning units. Table 4 summarises the eligibility criteria, patient characteristics, outcomes, and bed-day use of the cohorts eligible to be admitted to each of the four weaning units. Mortality was lower at ICU discharge and hospital discharge in unit A, the unit which accepted more stable patients (no RRT or vasoactive support for seven days prior to transfer).

The proportion of funded ICU bed-days used by these cohorts of patients varied from 8.1% to 9.7%, equivalent to a mean of 2.0 to 2.4 ICU beds occupied per day in the region over the two-year period. Figure 2 demonstrates how occupancy and refusal rates changed for weaning units A to D when capacity was varied from one to eight beds. For example, a three-bed unit with eligibility criteria for unit D would have a bed occupancy rate of 72.6%. However, a new admission would be refused admission on the day eligible for transfer to a weaning unit on 36% of occasions. A five-bed unit for unit D would have a bed occupancy of 48.6% and refusal rate for new admissions of 3%.

To take account of the effect of patients becoming unstable once transferred to a weaning unit, the modelling exercise was repeated. In this model, those patients who deteriorated and required cardiovascular support in units C and D, or renal and cardiovascular support in models A and B, were assumed to have been transferred out of the weaning unit and readmitted to the ICU on the first day of requiring organ support. For reasons of simplicity, once a patient was readmitted to the ICU, they were no longer eligible for transfer back to the weaning unit. The effect of this reduction in weaning unit bed occupancy and refusal rates is represented in Figure 2 using broken lines.

The potential cost saving gained from transferring patients to a weaning unit was modelled accounting for a limited bed capacity by varying unit capacity from one to eight beds (Figure 3). For unit D, a three-bed unit offers the highest cost saving of around £344,000 (€418,000) per year. Staffing a five-bed unit would no longer create a net saving. Instead, there would be a net cost of £35,000 (€43,000) per year. Once unstable patients who are transferred back to the ICU are taken into consideration, establishing unit D with four beds no longer has an associated cost saving (net cost £32,000 (€39,000) per year; Figure 3 broken line). The results of the sensitivity analysis, which was undertaken by varying the cost of a weaning unit bed from 50% to 100% of the cost of an ICU bed, are available in the Additional file 1 (Figures S1 and S2). It showed that once a weaning unit bed reached 70% of the cost of an ICU bed, a three-bed unit would no longer yield a cost saving for unit A.

The major strength of our study was the inclusion of 95% of all ICU admissions in the region over the study period, which minimised selection bias. Other studies have described a subpopulation of PMV patients following transfer to a weaning unit [15], or were unable to report characteristics of excluded patients [4]. We were also able to undertake a range of sensitivity analyses based on different definitions and inclusion of cohorts with uncertain PMV status. These analyses did not show clinically important variations in our estimates of incidence, suggesting they were accurate. Two groups of patients cared for in specific clinical areas were excluded from our study cohort because these areas did not submit data to the SICSAG database. Although most patients with neurological injury who required PMV were included in the study cohort, a small number who received ongoing invasive ventilatory support in the neurosurgical high dependency unit had this period of ventilation excluded. In addition, any patients ventilated in cardiothoracic ICU were excluded. Studies reporting PMV rates following cardiac surgery use a shorter period of ventilation to define PMV, making a comparison difficult with our findings. Two UK studies reported PMV rates of 5.5% (ventilation >48 hours) [16] and 2.6% (ventilation >96 hours) [17] in these patients. Therefore, although the patients are few in number, this could have resulted in a small underestimation of the overall regional incidence rate.

There are a number of limitations to our modelling exercise. The retrospective nature of our study meant that data relating to ventilatory modes, sedation practice, and weaning methods were not available. Advances in clinical practice, such as daily sedation breaks [18], spontaneous breathing trials [19], and weaning protocols [20] may help to reduce the length of MV and potentially reduce the need for establishing a weaning unit. We did not take the potential impact of a weaning unit on outcomes into account, for example, reduced duration of MV or improved mortality. A reduction in MV duration would further improve net cost savings and decrease patient refusals. In addition, we assumed patients would only be transferred after 21 days, whereas in practice transfer might occur at an earlier stage if appropriate, and according to bed pressures. These factors are difficult to adjust for in a modelling exercise, but in general would probably improve the efficiency of resource use. The method of costing health care in the UK is difficult to generalise to other health care systems, in part because costs are allocated using a 'top-down' approach rather than a per item 'bottom-up' approach [21]. Our estimates of cost included the assumption that both occupied and unoccupied weaning unit beds incurred similar costs, and we also did not consider the potential for transfer of suitable patients prior to 21 days of MV. These factors mean we have probably underestimated potential cost savings, although we did not include the capital costs associated with setting up a weaning unit. Despite this, our models are the first to use high-quality clinical data and consider different entry criteria, including adjustment for clinical deterioration. We also evaluated bed occupancy and patient refusal rates, which has not previously been undertaken.

Incidence estimates are influenced strongly by the denominator utilised, especially for critical care populations for which case-mix is dependent on health care organisation. The most externally valid denominator is probably patients who require MV in the ICU. Our data suggest a PMV incidence of 6.3 per 100 ventilated ICU patients in a typical UK health board region (or equivalent) for this population. Few studies of PMV incidence have been published and variations in both denominator and numerator definitions make direct comparison with our study difficult. Using the consensus definition of 21 days or longer of MV, a single-centre Argentinean study reported a PMV incidence of 14.3 per 100 ICU admissions [22], which is considerably greater than our cohort. Using a PMV definition of 21 days or longer of MV and a denominator of patients ventilated for 48 hours or longer, a prospective study based in a single US centre found that 14.0 per 100 patients required PMV [23]. This compares with a rate of 9.6 per 100 patients using similar definitions for numerator and denominator in our cohort (data not shown). A multicentre population based study in the US reported an incidence of PMV of 7.7 per 100 ventilated ICU admissions [24]. However, the numerator was defined as four days or longer ventilated with a tracheostomy and the denominator excluded elective surgical patients ventilated for less than 24 hours. Only one other UK study has reported PMV frequency. The incidence of stable PMV (≥14 days ventilated) patients who fulfilled criteria for a weaning unit was 2.5 per 100 ICU admissions [4]. Our data provide a number of measures of population incidence to improve generalisability of the findings.

The overall ICU and hospital mortality rates for the PMV cohort of 26% and 40% are consistent with severe critical illness. The higher death rate in the PMV cohort compared with the non-PMV cohort is likely related to the ongoing burden of chronic critical illness experienced by these patients [23]. Despite this, we found that 83% of PMV patients who were discharged alive from the ICU survived to hospital discharge. Direct comparison with published outcomes from other studies is difficult because most published cohorts had different entry criteria or were selected by admission to a weaning unit [25‐29].

A high proportion of patients were discharged to their own residence from the acute hospital, suggesting that rehabilitation is occurring primarily in the acute environment and/or that patients are discharged with significant ongoing rehabilitation requirements. A recent UK guideline has highlighted the need to improve rehabilitation following critical illness [13]. Establishing dedicated rehabilitation facilities may have the benefit of reducing hospital stay and improving patient outcomes.

Our modelling suggested that a three- or four-bed weaning unit was the most cost-efficient for our region. In practice it is likely that admission criteria might include patients from each of the four scenarios, with a key practical consideration being whether RRT in the form of intermittent haemodialysis would be provided in the unit. A geographically closer weaning unit would make provision of RRT or readmission to the ICU in the event of clinical deterioration more feasible. Only one other study has modelled a weaning unit in the UK. Robson et al. [4] used 14 days of ventilation rather than 21 days as the criterion for eligibility and required seven days of clinical stability prior to transfer. They found that 6.0% of ICU beds in their region were occupied by patients eligible for their weaning unit. Despite the limitations of modelling, our data suggest that a weaning unit could reduce the cost of providing critical care in our region if the equivalent number of acute ICU beds were reduced.

Further research is needed to evaluate the effect of establishing a weaning unit on patient-centred outcomes along with a more detailed economic evaluation. To date, published reports from US centres support possible improvements in some outcomes [30, 31], and a reduction in costs [31].