Effect of non-clinical inter-hospital critical care unit to unit transfer of critically ill patients: a propensity-matched cohort analysis

The Case Mix Programme (CMP) is the national clinical audit of adult, general critical care units (including intensive care and combined intensive care and high-dependency units) in England, Wales and Northern Ireland. The CMP is coordinated by the Intensive Care National Audit & Research Centre (ICNARC). Coverage of the database increased from over 60% of eligible units in 2008 to over 80% in 2010. Raw clinical data are abstracted retrospectively by trained local data collectors in accordance with precise rules and definitions. The data then undergo extensive validation, both locally and centrally. The CMP data collection and validation processes have been previously reported [12] and independently assessed to be of high quality [13]. CMP data collection covers the first 24-hour case mix (demographics, past medical history, surgical status, acute severity of illness and reason for admission) and outcomes (critical care unit and acute hospital discharge status). Support for the collection and use of patient-identifiable data without consent was obtained under Section 251 of the UK NHS Act 2006 (approval number: PIAG 2-10[f]/2005).

Patients aged 16 years and older were eligible for inclusion if they were admitted to an adult, general critical care unit in England or Wales that was participating in the CMP between 1 January 2008 and 30 September 2011. Subsequent critical care admissions to the same unit within the same hospital stay were excluded. Patients admitted to critical care units in Northern Ireland were also excluded because a comparable measure of socioeconomic position was not available.

Inter-hospital transfers were identified as patients discharged from the critical care unit to a level 3 bed in another acute hospital. Level 3 care is defined as the level of care for 'patients requiring advanced respiratory support alone or basic respiratory support together with support of at least two organ systems'. This level includes all complex patients requiring support for multiorgan failure [3]. Non-clinical transfers were defined as transfers for whom the reason for discharge from the original critical care unit was reported as transferred for comparable critical care (that is, for similar care as provided in the transferring unit). The CMP does not capture any additional information about the reason for transfer.

As acute severity of illness data are recorded in the first 24 hours following admission to a critical care unit in the CMP, patients transferred within 48 hours were selected to minimise the likelihood of acute severity of illness changing markedly in the time since assessment. The analysis was therefore restricted to patients who underwent a non-clinical transfer within 48 hours of admission to the original critical care unit.

A propensity model was built using logistic regression to model factors predictive of undergoing a non-clinical transfer, including patient and unit factors thought to be relevant on the basis of previous analyses. Patient factors included were: age; sex; deprivation; past medical history; admission type; surgical status; primary reason for admission to the critical care unit; acute severity of illness; month of admission to the critical care unit; and time of day of admission to the critical care unit.

Deprivation was assessed using area-based measures: the English Index of Multiple Deprivation (IMD) 2010 [14] and the Welsh IMD 2008 [15]. Each lower layer super output area (about 1,500 people) was assigned a deprivation score. Scores were then ranked and divided into quintiles (1 for least deprived, 5 for most deprived) and each patient was assigned to an IMD quintile based on their postcode of residence.

Past medical history was assessed against a list of specified, serious conditions, evident during the 6 months prior to admission to the critical care unit, and defined by the Acute Physiology and Chronic Health Evaluation II method [16]. Admission type was coded as per the Critical Care Minimum Data Set [17]. Surgical admissions were identified by admission to the critical care unit direct from theatre and recovery, and were subclassified as either elective/scheduled or emergency/urgent based on the National Confidential Enquiry into Perioperative Deaths classification [18].

The primary reason for admission to the critical care unit was coded using the ICNARC Coding Method [19] and was categorised by the underlying body system involved (for example, respiratory, neurological, and so forth). Acute severity of illness was assessed using the ICNARC Physiology Score, based on 12 physiological parameters measured during the first 24 hours in the critical care unit [20]. Unit factors included were the number of critical care beds in the transferring unit, and the critical care network of the transferring unit. Components of each of the variables are outlined in Table 1.

Continuous variables (age and ICNARC Physiology Score) were modelled using restricted cubic splines to allow for a flexible, nonlinear relationship between variables [21]. The model was then used to calculate a propensity of 0 to 1 for each eligible patient. Each patient undergoing a non-clinical transfer (case) was matched on the basis of absolute propensity (nearest-neighbour match) with two patients who were not transferred, but who had a length of stay at least as long as the time point at which the non-clinical transfer took place (controls). Sampling was carried out with replacement; patients thus may have acted as a control for more than one case.

To ensure that there was similarity between the two groups on important case-mix factors, the balance of patient-level factors between cases and matched controls was assessed using the chi-square test for categorical variables and using both quantile-quantile plots and the Kolmogorov-Smirnov test for continuous variables. The mean square difference in propensity between cases and controls was also calculated to assess the closeness of matching.

Comparisons of mortality were performed using conditional fixed-effects Poisson regression with 95% confidence intervals (CIs) estimated by bootstrap resampling with 500 replications (effect estimate: matched groups relative risk) [22]. Regression models were adjusted for the ICNARC model predicted log-odds of hospital mortality to further account for any residual differences in acute severity of illness at admission. In order to assess whether there was a significant change in acute severity of illness in the time between admission and transfer, we measured the difference in ICNARC Physiology Score in those cases in which it was possible to link data from their initial admission to critical care to information from their subsequent admission to a second unit following transfer. Comparisons of total acute hospital length of stay and total critical care unit length of stay were performed using repeated-measures analysis of variance.

The primary outcome was mortality at ultimate discharge from acute hospital (ultimate acute hospital mortality). Secondary outcomes were mortality at ultimate discharge from critical care (ultimate critical care unit mortality), total acute hospital length of stay and total critical care unit length of stay.

Statistical analyses were performed using Stata 10.1 (StataCorp LP, College Station, TX, USA).

During the study period, 308,323 eligible patients, aged 16 or over, were admitted to 198 participating adult general critical care units in England and Wales. These patients were admitted to critical care for the first time during that hospital stay. IMD quintiles could not be assigned to 2,578 (0.8%) patients and a further 11 (0.004%) were excluded because they were missing information about their source of admission. An additional 3,102 patients (1.0%) were excluded because they had missing information relating to one or more of the key outcome variables, whilst 1,577 patients (0.5%) were excluded because they had undergone a non-clinical transfer more than 48 hours after admission to critical care. We have provided additional data about the distribution of time from critical care unit admission to non-clinical transfer in the transferred population in Figure S1 of Additional File 1. Table S1 in Additional File 1 also presents information about the characteristics of patients undergoing a non-clinical transfer within 48 hours of admission to critical care, compared with after 48 hours. A propensity was calculated for 301,055 patients (97.6%), including 759 patients transferred for non-clinical reasons less than 48 hours after admission.

Table 1 and Figures 1 and 2 together show the key patient and unit-level factors that were predictive of a patient undergoing a non-clinical transfer, as demonstrated by the logistic regression analysis used to build the propensity model. Figures 1 and 2 illustrate the odds ratio of undergoing transfer by both age and physiology score relative to the mean of each variable as a baseline. The likelihood of transfer varies significantly with both age and physiology score. When the analysis was adjusted for all other factors, surgical patients, both emergency and elective, were significantly less likely to undergo a non-clinical transfer, as were those with endocrine or genitourinary (including renal) causes for admission. Patients admitted for neurological reasons were more likely to undergo a non-clinical transfer. In addition, the odds of patients undergoing a non-clinical transfer varied significantly throughout the year, being highest in January and lowest in July. In terms of unit-level factors, the odds of transfer decreased as the number of critical care beds in the transferring unit increased.

Each case was successfully matched with two controls. There was good balance between the two groups (Table 2 and Figures 3 and 4), with no significant differences in distributions. Only 11 controls (0.7%) were matched to two different cases and none were matched to more than two cases. The mean square difference in propensity between cases and controls was < 0.00001, giving assurance that the propensity matches were very close.

The crude acute hospital and critical care unit mortality were both slightly lower in the non-clinical transfer patients, compared with the propensity-matched controls (Table 3). In the propensity-matched analysis, adjusted for the ICNARC model predicted log-odds of hospital mortality, the relative risk of ultimate acute hospital mortality was 1.01 (95% CI = 0.87 to 1.16; P = 0.94) in patients who had undergone a non-clinical transfer within 48 hours following admission, compared with those who were not transferred. The relative risk of ultimate critical care unit mortality for nontransferred patients was 0.98 (95% CI = 0.82 to 1.18; P = 0.84).

The mean total length of stay in critical care was 3.2 days longer (95% CI = 2.1 to 4.3; P < 0.001) in the non-clinical transfer group. The mean total acute hospital length of stay was 3.0 days longer (95% CI = -0.1 to 6.1; P = 0.06).

Our analysis has several strengths when compared with the previous study [7]. First, we examined data on a large number of patients transferred for comparable critical care excluding those transferred for more specialised critical care. Second, given the high level of critical care unit participation, the CMP data are highly representative of patient experience in England and Wales. Third, the use of a propensity-based method enabled us to address the challenge of potential confounding that has hampered previous analyses. Fourth, we were able to closely match all 759 patients transferred for non-clinical reasons with two control patients sharing a similar propensity to undergo a non-clinical transfer. This approach enabled us to incorporate a greater number of possible confounders in the matching process than is feasible when using conventional methods.

Our analysis only included patients transferred between critical care units. However, we know that some patients are transferred directly from the emergency department of one hospital to a critical care unit in another hospital [24]. These patients are not captured by the CMP, but this group is likely to be sicker and less stable clinically. Further, quantifying the impact of non-clinical transfer in all patients, including those transferred from the emergency department, would require a large, prospective study in which acute severity of illness was recorded regularly so that cases could be matched based on the severity of their condition at the time of transfer. Given the relatively small number of non-clinical transfers that take place, this would need to be coordinated across a large number of centres, over many months, to ensure sufficient numbers of patients, making the study challenging and potentially unfeasible.

The CMP exists for national clinical audit purposes and only requires physiological data to calculate acute severity of illness scores in the 24 hours following admission to the critical care unit. We therefore restricted our analysis solely to those patients transferred for non-clinical reasons closer to this time period. However, omitting an unmeasured, yet potentially confounding, variable from the propensity model has been shown to result in biased estimation of the treatment effect [11]. To this end, there is a risk of residual confounding in our analysis associated with the timing of the measurement of acute severity of illness. Patients selected for non-clinical transfer may have a lower acute severity of illness at the time of transfer, compared with other patients in the critical care unit, and thus a lower risk of mortality. However, by restricting our analysis to non-clinical transfers within 48 hours of admission, there is less opportunity for the severity of illness to change markedly from the first 24 hours, so the risk of bias is minimised. In the group of transferred patients whose data could be linked between two critical care units, there was no statistically significant difference in acute severity of illness in the 24 hours following the second admission after transfer compared with the same period after initial admission. Another potential limitation is that patients transferred within 48 hours constitute only about one-third of all the non-clinical transfers that occurred and may therefore not be representative of all non-clinical transfers. Additionally, the difference in total critical care length of stay may be partly attributable to a selection bias; for example, if clinicians select a patient for non-clinical transfer who they consider is likely to require several more days of critical care.