Improving risk prediction model quality in the critically ill: data linkage study

The Case Mix Programme

The CMP is the national clinical audit for adult critical care with a remit for England, Wales and Northern Ireland. The CMP has been established for 20 years and the resulting high-quality clinical database (of > 2 million critical care admissions) has underpinned evaluations of policy and practice in critical care. 9 The CMP has 100% participation of adult, general critical care units delivering Level 3 or combined Level 2/3 care (intensive care units and combined intensive care/high-dependency units); over the time period of this study, participation of these units increased from 90% to approaching 100%. Participation of other critical care units, such as specialist units (e.g. neurocritical care units and cardiothoracic critical care units) and stand-alone Level 2 (high dependency) units is lower. Data on consecutive admissions to each participating critical care unit are recorded prospectively and abstracted from the medical records by trained data collectors according to precise rules and definitions. The data collected include demographics, past medical history, physiological and diagnostic data from the first 24 hours following admission to the critical care unit, outcome and activity data. The data undergo extensive validation checks, both within local software systems and centrally on submission to ICNARC, before being pooled into the CMP database. Details of data collection and validation have been reported previously, and the CMP database has been independently assessed and scored highly by the Directory of Clinical Databases (DoCDat) against their 10 domains (describing elements of coverage and accuracy). 3

The National Cardiac Arrest Audit

The NCAA was established in 2009 as a joint venture between ICNARC and the Resuscitation Council (London, UK). The NCAA is the national clinical audit of patients aged > 28 days in acute hospitals in the UK who receive cardiopulmonary resuscitation (CPR) and are attended by the hospital-based resuscitation team (or equivalent) in response to a 2222 call (2222 is the emergency telephone number used to summon a resuscitation team in UK hospitals). CPR is defined in NCAA as the receipt of chest compressions and/or defibrillation. Standardised data are collected at the time of the cardiac arrest and from the medical records according to precise rules and definitions. Staff members at participating hospitals enter data onto a dedicated secure online data entry system. Data are validated, both at the point of entry and centrally, for completeness, logicality and consistency. Details of data collection and validation have been reported previously. 4

The UK Renal Registry

The UKRR (https://renal.org/about-us/who-we-are/uk-renal-registry; accessed 28 October 2022) was established by the Renal Association (now the UK Kidney Association; Bristol, UK) and provides a focus for the collection and analysis of standardised data relating to the incidence, clinical management and outcome of end-stage renal disease. The Registry has been in operation since 1995, with 100% coverage of adult renal units in England and Wales since 2007.

The National Diabetes Audit

The NDA (https://digital.nhs.uk/data-and-information/clinical-audits-and-registries/national-diabetes-audit; accessed 28 October 2022) is the largest annual clinical audit in the world and is managed by NHS Digital working with Diabetes UK (London, UK) and the National Cardiovascular Intelligence Network, Public Health England. The National Diabetes Core audit, covering care processes, treatment targets, complications and mortality for people with diabetes in primary care and specialist services, is now in its ninth year. Over recent years, the audit has included > 80% of people diagnosed with diabetes in England and Wales.

The National Adult Cardiac Surgery Audit

The NACSA (www.nicor.org.uk/national-cardiac-audit-programme/adult-cardiac-surgery-adult-surgery-audit/; accessed 28 October 2022) collects consecutive operation data from all 35 NHS hospitals in the UK that carry out adult heart surgery. It has been running since 1977, making it the longest running of all UK national clinical audits. The audit is managed by the National Institute for Cardiovascular Outcomes Research (NICOR) at Barts Health NHS Trust, in association with the Society for Cardiothoracic Surgery.

Hospital Episode Statistics for England

The HES database is produced by NHS Digital using records provided by hospitals for the purpose of reimbursement of services funded through the NHS. 10 This study used the admitted patient care (APC) section of the HES database, which contains one record for each episode of care under one consultant during a hospital admission. The HES data set captures all publicly funded hospital activity, which is estimated to represent 98–99% of all hospital activity in England. 10 APC records include information about diagnoses and treatments received, alongside admission details (provider, dates, locations, etc.) and limited patient demographics.

Office for National Statistics death registrations

The ONS death registrations (mortality database) contains information abstracted from civil registrations of deaths registered in England and Wales. The database captures 100% of deaths registered in England and Wales but may not capture the deaths of English/Welsh residents occurring in other countries. Registration is delayed in cases of coroners’ investigations. The database includes information about the date, cause and location of death. Although constructed by the ONS, the database is routinely linked to HES and available jointly through NHS Digital.

Case Mix Programme cohorts

For the CMP admission cohort, we selected patients admitted to NHS adult critical care units in England participating in the CMP with identifiable linkage with HES and death registrations. False linkage and errors in linkage with HES and death registrations were excluded. The linked cohort includes patients admitted to participating critical care units between 1 April 2009 and 31 March 2015. The final follow-up date for death registrations was 15 March 2015. For patients with multiple hospital episodes that included critical care unit admissions, we considered the index hospital admission to be the first hospital admission during the analysis period. Re-admissions to critical care and transfers between critical care units during the index hospital admission were excluded.

For the CMP hospital survivor cohort, we selected from the CMP admission cohort those patients who survived to discharge from acute hospital (i.e. the end of the index hospital admission).

For the CMP cardiothoracic critical care cohort, we selected from the CMP admission cohort those patients who were admitted to a cardiothoracic critical care unit participating in the CMP.

National Cardiac Arrest Audit cohorts

For NCAA, data are collected for all individuals (excluding neonates) receiving chest compressions and/or defibrillation and attended by a hospital-based resuscitation team (or equivalent) in response to a 2222 call (2222 is the telephone number used to summon a resuscitation team in UK NHS hospitals). For the NCAA in-hospital cardiac arrest cohort, we selected validated team visits from hospitals in England participating in NCAA with valid linkage with HES and death registrations between 1 October 2009 and 31 March 2015. False linkage and errors in linkage with HES and death registrations were excluded. The final follow-up date for death registrations was 15 March 2015. For patients with multiple hospital episodes that included an in-hospital cardiac arrest, we considered the index hospital admission to be the first hospital admission during the analysis period. Individual team visit records meeting the following criteria were excluded from the cohort: arrests that occurred pre hospital (but were subsequently attended by a hospital-based resuscitation team – usually in the emergency department – and therefore met the scope of NCAA); second and subsequent visits to the same patient during the same hospital stay; and patients for whom it was identified, after commencing resuscitation, that a ‘do not attempt cardiopulmonary resuscitation’ decision was already documented in the patient’s notes.

For the NCAA critical care admission cohort, we selected from the NCAA in-hospital cardiac arrest cohort those patients with valid linkage with CMP, for whom a critical care unit admission occurred after the first in-hospital cardiac arrest recorded in NCAA and during the index hospital admission.

Mortality at discharge from acute hospital

Outcome defined as mortality at discharge from acute hospital (acute hospital mortality), as used for the current risk models. Patients transferred to another acute hospital were followed up until final discharge.

Mortality at 30 days, 90 days and 1 year and survival time

Outcome was defined as mortality at 30 days, 90 days and 1 year following critical care unit admission and time from critical care unit admission or discharge alive to death using the date of death obtained by data linkage with death registrations. In addition, 1-year post-discharge mortality was defined as mortality at 1 year following discharge from acute hospital for patients who survived the index hospitalisation.

Hospital resource use and costs post-critical care

Outcome was defined as number of days in acute hospital, either during the original hospital episode (as identified from CMP data) or during subsequent hospital episodes (as identified through data linkage with HES) and total hospitalisation costs calculated from NHS reference costs. 15

New diagnosis of end-stage renal disease post-critical care

Outcome was defined as new receipt of RRT for ESRD, based on the date of diagnosis recorded in the UKRR database, after the date of discharge from hospital.

New diagnosis of diabetes post-critical care

Outcome was defined as a new registration for type 2 diabetes, based on the date of diagnosis recorded in the NDA database, after the date of discharge from hospital.

Return of spontaneous circulation > 20 minutes (National Cardiac Arrest Audit)

Outcome was defined as return of spontaneous circulation (ROSC) sustained for > 20 minutes.

Hospital survival (National Cardiac Arrest Audit)

Outcome was defined as survival to discharge from the hospital in which the in-hospital cardiac arrest occurred. Patients transferred to another acute hospital are counted as hospital survivors.

Comorbidities

There are two measures to derive comorbidities: severe conditions in the past medical history – defined according to the Acute Physiology And Chronic Health Evaluation (APACHE) II method18 (Table 2) – and comorbidities based on Armitage and van der Meulen’s 2010 Royal College of Surgeons (RCS) Charlson Score17 (Table 3). Severe conditions in the past medical history are recorded in the CMP. Conditions must have been evident in the 6 months prior to admission to the critical care unit and documented in the patient’s notes at or prior to admission. RCS Charlson comorbidities were identified from linked HES records by relevant ICD-10 codes (see Table 3) in any of the first seven diagnosis fields of the index hospital admission or of any episode that finished during the year preceding the index hospital admission.

To allow the use of both severe and lower levels of comorbidity in the same model, and reduce duplication of these measures, we combined the nine APACHE II comorbidities and 13 of the RCS Charlson comorbidities to produce 19 comorbid categories (Table 4). Acquired immune deficiency syndrome (AIDS) and human immunodeficiency virus (HIV) were not included because these diagnosis codes were supressed in the HES extract, and CMP data indicate it was a rare diagnosis in this patient group.

Sepsis

The CMP database can be used to identify patients who had sepsis at admission to the unit or who developed sepsis during the first 24 hours in the unit based on their reason for admission to critical care and physiology measured during the first 24 hours following admission. It cannot be used to identify patients who developed sepsis after the first 24 hours following admission to the unit.

Based on the current international consensus definitions,20 sepsis-3 is defined as infection plus new organ dysfunction, defined as an increase in a Sequential Organ Failure Assessment (SOFA) score of ≥ 2 or points. In the CMP, this is operationalised as evidence of infection from the primary or secondary reason for admission to the critical care unit plus organ dysfunction, defined as a SOFA score of ≥ 2 points in any one organ system or a SOFA score of ≥ 1 points in two or more organ systems, based on physiological data from the first 24 hours following admission. 21 Full details of the organ dysfunction definitions are summarised in Appendix 1, Table 37.

Handling of missing data

The percentage of physiological predictors with missing values in the CMP data set ranged from 0.6% for highest heart rate to 13.9% for blood lactate. In a previous work6 exploring the impact of missing data on developing ICNARC risk prediction models using CMP data, no differences in the inference were found and coefficient estimates appeared to be insensitive to the missing data and the various models used to deal with them. The study6 concluded that under the CMP, missingness scenario benefits of using multiple imputation in developing the risk prediction model are likely to be minimal. Therefore, for objective 1 (risk models for adult critical care using the full CMP cohorts) we decided that the model building and analysis process would be done with non-imputed (complete-case) data and a parallel analysis would be done at the same time on the multiply imputed data set to test the consistency of the results.

For analysis relating to objective 2 (risk models for cardiothoracic critical care) and objective 3 (risk models for in-hospital cardiac arrest), missing data were imputed to address potential bias and loss of precision in these smaller data sets. Fully conditional specification was used as the multiple imputation method. 22 All the candidate predictors (with or without missing values) and the outcome, as well as auxiliary variables related to missingness, were entered into the imputation model. 23,24 When required, simple or zero-skewness log transformation for non-normality was applied. Unless the rate of missing information is unusually high, there tends to be little or no practical benefit to using more than 10 imputations and so, in the following analysis, 10 repeat imputations were performed.

Patient characteristics

Demographic, clinical and other characteristics of the patients included in the development and validation data sets for the risk prediction models for each objective were summarised. Categorical variables were summarised by frequencies and percentages. Percentages were calculated according to the number of patients for whom data were available. Continuous variables were summarised by mean and standard deviation (SD), as well as median and interquartile range (IQR).

Analysis of outcome measures

Methods for model development were based on those used for the development of risk models for acute hospital mortality in the previous NIHR study. 6 Binary outcomes were modelled with logistic regression models.

Models for longer-term chronic health outcomes analyses had to account for competing events, in particular when some patients die before being observed at the stage of interest. Death should be considered as a competing event because it precludes the observation of the stage of interest. To account for both the time-to-onset and competition with death, cause-specific Cox proportional hazards models25 and Fine/Gray models26 were considered.

It is well recognised that the statistical analysis of health-care resource use and cost data poses a number of difficulties. First, we reviewed the literature on methods and models in the field of calculating and predicting health-care resource use and costs, their ability to address with the usually statistical issues that is, skewness, excess zeros, multimodality and heavy right tails, as well as specific challenges of the project such as competing risk and censoring. Before a decision was made, we met with health economics experts to discuss the above issues. Finally, a generalised linear model with a gamma distribution and a log-link function was agreed as the most reasonable approach. 27

An analysis plan was finalised with input from the Clinical Advisory Group for each objective prior to modelling.

In all, the following approaches for model development were applied depending on the outcome and objectives of the analysis:

• To model mortality/survival at fixed time points (hospital discharge, 30 days, 90 days, 1 year, ROSC > 20 minutes): logistic regression (including, if appropriate, random effects of critical care unit/hospital).

• To model time-to-event outcomes: standard survival regression methods such as Kaplan–Meier survival curves and Cox regression.

• To model chronic health outcomes with a competing risk of death: cause-specific Cox proportional hazards models and Fine/Gray models.

• To model critical care/hospital resource use and costs: generalised linear model with a gamma distribution and a log-link function.

Functional form

In the previous project,6 the functional form of physiological predictors and the optimal approach to deal with these were explored. We decided to use restricted cubic splines because they showed the flexibility of fractional polynomials but with better behaviour in the tails, they captured the most prominent features of the relationship between predictors and outcomes, and the fit was more plausible than the previous categorical approach.

So, after having rejected the hypothesis of linearity, the following approaches for modelling continuous predictors were applied.

The predictor–outcome relationship was explored by expanding the variable into multiple terms and testing pooled and individual non-linearity. Spline fits could be sensitive to the number of knots so, to avoid overfitting, spurious dips and inflexion points, as well as unrealistic features of the curve, three, four or five knots were considered. Knot positions were selected according to the recommendations of Harrell. 28 Right restricted cubic splines (i.e. with the linearity restriction applied only at the right-hand end of the curve) were used when appropriate (e.g. for variables that were bounded by zero) to allow more flexibility.

To judge the plausibility and accuracy of the fitted curves, we plotted observed log-odds against the alternative modelling approaches and used a running line smoother as a reference. The Akaike information criterion (AIC) and Bayesian information criterion (BIC) were calculated to compare the fit of the strategies for modelling continuous variables, taking three-knot restricted cubic splines as reference, to assess if the increased complexity of the model resulting from including more knots was worthwhile.

The best functional form for each predictor was selected based on fitting, plausibility, accuracy and prior knowledge about the predictor.

We finally explored collapsing extreme points to determine if the shape of the curves could be affected by outlying values. Additional analyses using imputed data were done in parallel to provide reassurance about the results.

Approach to model development

For each newly developed or revised risk prediction model, we adopted the following general approach to model development:

1. Potential predictors were identified and patterns of missing data within the potential predictors were explored, with particular attention to the completeness and accuracy of data linkage between the databases. Approaches to handling the missing data were compared, based on the best performing approaches from previous work.

2. The most appropriate functional form for each potential predictor was explored, taking into consideration the use of continuous non-linear models (e.g. restricted cubic splines or right-restricted cubic splines) for continuous predictors and appropriate categorisation and structure of categorical predictors.

3. A main effects model was fitted through a process of deleting terms, re-fitting and verifying, using Wald and/or likelihood ratio tests to remove non-significant predictors, and the BIC as the basis to determine which predictors make an important contribution to the fit of the model.

4. The functional form and significance of each predictor included in the main effects model were then re-examined to confirm if any changes were required based on adjustment for other important predictors.

5. Finally, interactions between the predictors were introduced based on clinical input to identify and prioritise the potentially important interactions to consider and avoid over fitting, with interactions retained if they have a positive effect on the BIC.

The starting point for revised risk models was the previously developed risk model for each outcome. The addition of new predictors was considered and then the effect of those predictors previously included in the existing models was re-assessed to determine whether or not they still made an important contribution to the model (see point 3 above).

For developing risk models for mortality/survival at fixed time points and time-to-event outcomes (objectives 1a, 2b, 3b), the starting point for each new risk model was the risk model for survival to hospital discharge developed in the previous objective (2a and 3a) and the existing model for adult general critical care. In each case, this risk model was refitted to the new outcome, risk factors that were previously considered but were found not to be important predictors for hospital survival were reassessed by adding them to the model and, finally, the effect of those risk factors previously included in the risk model was reassessed to determine whether or not they still made an important contribution to the model (see above). It was anticipated that predictors representing age, chronic ill health and functional status would have a greater impact on longer-term outcomes than on hospital survival, whereas predictors relating to the acute illness would have less impact.

As the important risk factors and their relationships with the outcome may be very different from those considered previously, the new risk models for critical care resource use and for longer-term chronic health outcomes were developed, de novo, using the methods previously described.

Assessing the predictive performance

Throughout the study, risk prediction models were validated for their discrimination, calibration and overall fit. The panel of measures described here was used to give an overall assessment of model performance.

The discrimination of the model was estimated by the c index (equivalent to the area under the receiver operating characteristic curve)29 and accuracy was assessed by Brier’s score (mean squared error between outcome and prediction). 30 We assessed calibration graphically with predicted probability on the x-axis and the observed outcomes on the y-axis in 10 equal-sized risk groups (calibration plot) and by Cox’s calibration regression (linear recalibration of the predicted log odds). 31 The standardised mortality ratio (SMR) with 95% confidence interval (CI) was calculated to observe the difference between actual and expected mortality by dividing the observed number of deaths by the number of deaths predicted by the model. Because of the size of the data sets to be used, we did not assess the calibration of the model with the Hosmer–Lemeshow c-statistic because this may have led to misleading conclusions. 32

Internal and external validation

Each newly developed or revised risk prediction model was validated using the above measures both within the development sample and in independent validation data.

When a prognostic model is based on a very large sample size and relevant variables are included in the final model, optimism is small and so, the apparent estimates of model performance (c index and Brier’s score in the development data) are attractive because of their stability. 33 However, to assess optimistic performance within the development data, the percentage of over-fitting was estimated by refitting the model in 100 bootstrap replications of the data set and evaluating the resulting model in the original development data to calculate the optimism-corrected statistics. 34,35

When existing risk prediction models were modified, the performance of the revised model was compared with the existing model using reclassification techniques. 36 The improvement in reclassification was quantified as the net reclassification improvement (NRI). NRI is an index that attempts to quantify how well a new model reclassifies subjects – either appropriately or inappropriately – as compared with an existing model. We calculated the NRI both using pre-defined categories of risk and also as a continuous measure (i.e. the proportion with any improvement in predicted risk compared against the proportion with any worsening in predicted risk).

In addition, the risk models developed were compared, when relevant, against existing risk models (e.g. the ICNARC_H–2015 model7).

Time-to-event analysis

Statistical methods relevant to model time-to-event outcomes are described in the corresponding chapters.

Study cohort

The cohort for this chapter was the CMP hospital survivor cohort (see Chapter 3).

Inclusion and exclusion criteria

From the CMP hospital survivor cohort, patients were selected if they were discharged from acute hospital (from their index hospital admission) between 1 April 2009 and 15 March 2010 – allowing 5 years of follow-up for all patients to the final follow-up date for death registrations of 15 March 2015.

Outcome measure

The outcome measure was time to death following discharge from acute hospital. The outcome was defined as death (from any cause) within 5 years of discharge from hospital and number of days from hospital discharge to death, established by data linkage with death registrations. Patients not reported to have died were censored at 15 March 2015.

Statistical analysis

Kaplan–Meier survival curves were generated to describe mortality after hospital discharge among critical care admissions. Estimated cumulative mortality from 30 days to 5 years is reported both overall and for patient subgroups identified from the primary reason for admission to the critical care unit (pneumonia, bowel tumour, traumatic brain injury, cardiac surgery). The survival curve for critical care survivors was compared with that for the age- and sex-matched general population.

Outcomes and candidate predictors

Risk prediction models were developed for three outcomes: mortality at 30 days, 90 days and 1 year following admission to the critical care unit. Patients were followed up to each end point from the first critical care unit admission during the index hospital admission. This starting point was selected for a benchmarking purpose. 43 A second 1-year follow-up time point from hospital discharge for hospital survivors was also established with the aim to compare the role of comorbidities among hospital survivors.

The starting point for the risk prediction models in the present chapter was the set of physiological and non-physiological predictors from the CMP previously included in the existing model to predict mortality at discharge from acute hospital. 7 Severe conditions in the medical history that were previously considered but were found not to be important predictors for acute hospital mortality were reassessed for short-term outcomes. The full list of candidate predictors from the CMP is presented in Appendix 1, Table 38.

It was anticipated that chronic ill health has had a greater impact on longer term outcomes than on mortality at 30 days or at discharge from acute hospital. Combinations of severe conditions in the past medical history (APACHE II) and RCS Charlson comorbidities, as described in Chapter 3, were therefore included in the 1-year analyses.

Development of a risk model for 30-day mortality

The steps to developing a risk model for 30-day mortality, starting from the predictors in the model for acute hospital mortality, were as outlined in Chapter 3.

Reasons for admission to critical care are recorded in the CMP using the ICNARC Coding Method: a five-tiered (type – surgical or non-surgical/body system/anatomical site/physiological or pathological process/condition) coding system specifically developed for this purpose. 44 Currently, coefficients for the ICNARC model are applied at three levels of the hierarchical code—either at tier 5, the individual condition, or at tier 4, the process, or at process/the body system combination. The following steps were followed in updating the primary reason for admission categories:

• As starting point, the current process and process/system categories were retained.

• The set of categories was refined by adding specific conditions.

• Process/system categories were split into individual conditions that had sufficient sample size (number of events ≥ 20).

• Each individual condition was retained as a new category if it was significant as a stand-alone variable in the model after adjusting for process/system (likelihood ratio tests, p < 0.001) and made an important contribution to the fit of the model (based on the presence of a strong effect on the BIC).

Finally, interactions between the categories and physiology were introduced based on clinical input to identify and prioritise the potentially important interactions to consider and avoid overfitting, with interactions retained if their likelihood ratio test p-value was < 0.001 and had a positive effect on the BIC.

Use of 30-day mortality for benchmarking

Existing risk prediction models for critical care, including the ICNARC model, have been based on an outcome of mortality at hospital discharge. These models have been used for national clinical audits, including the CMP and NCAA, to underpin fair comparisons among health-care providers. Research from the Netherlands42 has suggested that comparison of risk-adjusted mortality across critical care units using mortality at 30 or 90 days, rather than at hospital discharge, results in less heterogeneity.

To assess the effect of using 30-day mortality for benchmarking instead of acute hospital mortality, the final model for 30-day mortality was used to predict acute hospital mortality in the development data set. We used second-level customisation: acute hospital mortality was set as the outcome and the predictors were the variables included in the final model for 30-day mortality. The performance of this customised model (discrimination, calibration and accuracy) was used to assess whether or not the 30-day model could be used to predict mortality at acute hospital discharge.

As the SMR is used on benchmarking to evaluate the performance of a critical care unit we explored the effect of using 30-day versus in-hospital mortality on the SMR and the impact of the location of critical care units on a funnel plot of SMR against sample size. 45 Upper two and three SD control limits were used to identify higher-than-expected mortality and lower two and three SD control limits were used to identify lower-than-expected mortality.

Development of a risk model for 90-day mortality

The same approach used to customise the model to predict acute hospital mortality was applied when extending mortality prediction from 30 days to 90 days. A second-level customisation of the final 30-day mortality model was used to predict mortality at 90 days following admission to critical care.

Development of a risk model for mortality at 1 year following critical care admission

The risk model for mortality at 1 year following critical care admission was developed using the methods outlined in Chapter 3 using, as a starting point, the previous model for acute hospital mortality and considering the additional comorbidity variables in the development.

Development of a risk model for mortality at 1 year following hospital discharge

The risk model for mortality at 1 year following hospital discharge was developed using the methods outlined in Chapter 3 using, as a starting point, the previous model for mortality at 1 year following critical care admission.

Mortality at 30 days

Between 1 January 2014 and 31 December 2014, there were 153,494 first admissions to 235 adult critical care units in England participating in the CMP with identifiable linkage with HES and death registrations, of which 123,719 (80.6%) had complete data for all candidate physiological predictors (see below). Patients who were dead, were in line for palliative care or had all active treatment withdrawn immediately on admission and patients with missing values in non-physiological predictors were excluded. In total, 119,509 patients were included in the development data set. As described in Chapter 3, missing values were not imputed in the primary analyses for adult critical care.

Of the 119,509 patients, 22,579 (18.9%) died during the 30-day follow-up. Of these, 16,129 (13.5%) died during the critical care admission. The median critical care unit length of stay was 57 hours (IQR 26–122 hours) for 30-day survivors and 68 hours (IQR 28–151 hours) for 30-day non-survivors. The median hospital length of stay was 14 days (IQR 7–29 days) for 30-day survivors and 7 days (IQR 3–15 days) for 30-day non-survivors.

The characteristics of the included patients are described in Table 9. The median age was 66 years (IQR 52–75 years) and most of the admissions were able to live without assistance in daily activities (76.4%).

The most common severe condition in the medical history was immunocompromise (7.3%). Regarding RCS Charlson comorbidities, chronic pulmonary disease (13.2%), diabetes mellitus (13%) and malignancy (10.7%) were the most prevalent.

Model development

Functional form and significance of CMP physiological and non-physiological predictors previously included in the existing model for adult general critical were reassessed. Consistency with both optimal functional form and global significance of the predictors was found after fitting a model considering all physiological and non-physiological predictors from the current ICNARC model (Table 10).

TABLE 10 - Model performance in the development cohort

LL, log-likelihood.

Physiological and non-physiological predictors from the CMP previously included in the existing model to predict mortality at discharge from acute hospital.

Model	df	LL	BIC	C index (95% CI)	Brier’s score
Main modela	110	–36784.79	74867	0.888 (0.886 to 0.890)	0.095
Main model + new predictors	115	–36641.15	74638	0.889 (0.887 to 0.891)	0.095
Main model + new predictors + reason for admission	184	–35210.77	72584	0.898 (0.896 to 0.901)	0.091
Final model	226	–35028.98	72711	0.900 (0.897 to 0.902)	0.091

In line with the previous project,6 we decided to use restricted cubic splines to model continuous candidates because these showed the flexibility of fractional polynomials but with better behaviour in the tails and captured the most prominent features of the relationship between predictors and outcomes, and because the fit was more plausible than a categorical approach. 6 For body mass index (BMI), a simplification using three knots was enough to accommodate the non-linear behaviour and had better AIC and BIC than four knots.

Severe conditions in the medical history were added into the model, but the only conditions finally retained were very severe cardiovascular disease, severe respiratory disease and immunocompromise, because of their significant effect and contribution to the model.

After adjusting for current and new potential predictors, deprivation was not retained in the main term model owing to the lower contribution to the model fit.

The new categories of primary reason for admission are shown in Table 11. Conditions were selected after the modelling process described above. A total of 57 process/system combinations and 13 individual conditions from the ICNARC Coding Method form the new reason for admission categories. These accounted for 94.5% and 5.5% of admissions, respectively. Fifteen significant primary reasons for admission/physiology interactions were retained in the model.

TABLE 11 - Final categories of reason for admission included in the model

Reason of admission categorical variable	Frequency	Percentage
Combinations of process and system
Accidental intoxication or poisoning (endocrine)	398	0.33
Acidaemia (endocrine)	336	0.28
Burns or hyperthermia (dermatological)	115	0.10
Collapse (respiratory)	501	0.42
Coma or encephalopathy (neurological)	991	0.83
Congenital or acquired deformity or abnormality
Cardiovascular	556	0.47
Cardiovascular, endocrine, gastrointestinal, genitourinary or haematological/immunological	1549	1.30
Musculoskeletal	1175	0.98
Neurological	654	0.55
Respiratory	763	0.64
Degeneration
Cardiovascular	2442	2.04
Neurological	39	0.03
Diabetes mellitus (endocrine)	1790	1.50
Dissection or aneurysm (cardiovascular)	3633	3.04
Failure
Cardiovascular	1138	0.95
Genitourinary	4616	3.86
Haemolysis or thrombocytopaenia	93	0.08
Haemorrhage
Cardiovascular	141	0.12
Gastrointestinal	2374	1.99
Genitourinary	660	0.55
Neurological	1709	1.43
Respiratory	151	0.13
Hyperkalaemia (endocrine)	284	0.24
Hypokalaemia (endocrine)	103	0.09
Hyponatraemia (endocrine)	266	0.22
Hypoplasia or dysplasia (haematological/immunological)	121	0.10
Hypothermia (endocrine)	83	0.07
Infection
Cardiovascular	451	0.38
Dermatological, gastrointestinal, haematological/immunological, musculoskeletal or neurological	6438	5.39
Genitourinary	1878	1.57
Respiratory	6133	5.13
Inflammation
Cardiovascular, dermatological, genitourinary, musculoskeletal, respiratory	2863	2.40
Gastrointestinal	10,189	8.53
Neurological	177	0.15
Non-traumatic aneurysm, dissection, perforation or rupture (cardiovascular)	2677	2.24
Obstruction
Cardiovascular	7948	6.65
Gastrointestinal	4288	3.59
Genitourinary	606	0.51
Respiratory	1255	1.05
Oedema, inflammation, fibrosis or inhalation (respiratory)	104	0.09
Seizures (neurological)	2388	2.00
Self-harm or self-poisoning (endocrine)	2950	2.47
Shock and hypotension (cardiovascular)	2003	1.68
Transplant or related (cardiovascular, endocrine, genitourinary, haematological/immunological, respiratory)	297	0.25
Transplant or related (gastrointestinal)	523	0.44
Trauma
Neurological	391	0.33
Perforation or rupture (cardiovascular)	334	0.28
Perforation or rupture (dermatological, genitourinary, musculoskeletal, respiratory)	3348	2.80
Perforation or rupture (gastrointestinal)	5666	4.74
Perforation or rupture (neurological)	2014	1.69
Tumour or malignancy
Cardiovascular, dermatological, endocrine, gastrointestinal, musculoskeletal, respiratory	12,089	10.12
Genitourinary	4065	3.40
Haematological/immunological	414	0.35
Neurological	2117	1.77
Vascular
Cardiovascular; genitourinary	138	0.12
Gastrointestinal	891	0.75
Neurological	1579	1.32
Specific conditions
Acute alcoholic hepatitis	185	0.15
Alcoholic cirrhosis	325	0.27
Anoxic or ischaemic coma or encephalopathy	1037	0.87
Asthma attack in new or known asthmatic	1147	0.96
Fungal or yeast pneumonia	148	0.12
Hanging or strangulation	153	0.13
Intracerebral haemorrhage	1178	0.99
Multiple rib fractures	276	0.23
Non-traumatic subdural haemorrhage	390	0.33
Pulmonary fibrosis or fibrosing alveolitis	182	0.15
Secondary hepatic tumour	895	0.75
Secondary hydrocephalus	137	0.11
Thrombo-occlusive disease of brain	561	0.47

Interactions included in the previous model between severe liver disease with physiology, CPR with physiology, ventilation with physiology and physiology with physiology were re-assessed. Only the interactions between severe liver disease and temperature and between ventilation and heart rate were not significant after adjusting for the main term model plus the new primary reason categories and interactions. They were finally dropped.

Following the development process, the significance and importance of the predictors in the final model are shown in Appendix 2, Table 41. Full coefficients for the final model are presented in Appendix 3, Table 45. The distribution of predicted acute hospital mortality from the new model is shown in Figure 3. The final model showed good performance (c index 0.900 and Brier’s score 0.091) and, internally, calibration of the model was satisfactory (Figure 4). Overfitting was of limited relevance because of the very large data set and, as expected, model optimism was negligible (0.49% estimated overfitting).

FIGURE 3.

Distribution of predicted risk from the final risk model for mortality at 30 days following critical care admission in the development data set.

FIGURE 4.

Calibration of the final risk model for mortality at 30 days following critical care admission in the development data set.

The estimates for the model parameters obtained using data from the multiply imputed data set were similar to values estimated from the development data set and, therefore, the bias that could arise from using only the available information was considered to be very small.

Model validation

The performance in the validation data set of 134,750 admissions from 31 March 2015 to 31 March 2016 is presented in Table 12 and Figure 5. Discrimination and accuracy were excellent: a c index of 0.90 (95% CI 0.89 to 0.91) and Brier’s score of 0.088. The calibration of the model was satisfactory, with a calibration slope of 0.998 and a calibration intercept of 0.013. The observed and mean predicted mortality were 17.8.% and 17.6%, respectively, for a SMR (observed divided by predicted mortality) of 1.01 (95% CI 1.00 to 1.02), which indicated a good calibration in the main.

TABLE 12 - Overall predictive performance of the final risk model for mortality at 30 days and customised risk model for mortality at 90 days following critical care admission in the external validation data set

Model	c index (95% CI)	Brier’s score	Predicted mortality (%)	Observed mortality (%)	SMR (95% CI)
Final risk model for mortality at 30 days	0.900 (0.897 to 0.912)	0.088	17.69	17.82	1.01 (1.00 to 1.02)
Final customised risk model for mortality at 90 days	0.883 (0.880 to 0.884)	0.105	21.47	21.53	1.00 (0.99 to 1.01)

FIGURE 5.

Calibration of (a) the final risk model for mortality at 30 days following critical care admission and (b) the customised risk model for mortality at 90 days following critical care admission in the external validation data set.

Use of 30-day mortality for benchmarking

The performance in the development data set for the customised model for predicting acute hospital mortality was excellent: a c index of 0.899 (95% CI 0.897 to 0.901) and Brier’s score of 0.095. The calibration of the model was satisfactory (Figure 6). This indicated that the model performs well when applied for prediction of acute hospital mortality.

FIGURE 6.

Calibration of the customised risk model for acute hospital mortality in the development data set.

The effect of benchmarking by using mortality at 30 days after critical care admission instead of acute hospital mortality was assessed in the 232 included critical care units. There was strong agreement between SMRs calculated on 30-day mortality and acute hospital mortality (Supplementary material, Figure S2). Funnels plots of SMR for individual critical care units (Figure 7) show that 21 units (9%) moved across the 2SD (dashed funnel line) and 3SD (solid funnel line) boundaries when the outcome was changed to 30-day mortality. Table 13 summarises the comparison of SMR position between acute hospital mortality and 30-day mortality. When changing from acute hospital mortality to 30-day mortality, 12 units moved to less extreme positions in the funnel (one from below 3SD to inside the funnel, six from below 2SD to inside the funnel and five from above 2SD to inside the funnel) and nine units moved to more extreme positions in the funnel (one from below 2SD to below 3SD, five from inside the funnel to below 2SD, one from inside the funnel to above 2SD and two from above 2SD to above 3SD). This suggests that there was little difference in heterogeneity across critical care units when assessed with 30-day mortality rather than acute hospital mortality.

FIGURE 7.

Funnel plots of SMR for (a) acute hospital mortality and (b) 30-day mortality.

TABLE 13 - Comparison of SMR positions in the funnel plots based on acute hospital mortality versus 30-day mortality

Shaded cells signify no change in position within the funnel plots.

Acute hospital mortality	30-day mortality
	Below 3SD	Below 2SD	Inside funnel	Above 2SD	Above 3SD	Total
Below 3SD	3	0	1	0	0	4
Below 2SD	1	2	6	0	0	9
Inside funnel	0	5	199	1	0	205
Above 2SD	0	0	5	4	2	11
Above 3SD	0	0	0	0	3	3
Total	4	7	211	5	5	232

Mortality at 90 days

The performance of the customised model for 90-day mortality was good, presenting a c index of 0.883 (95% CI 0.881 to 0. 885) and a Brier’s score of 0.1077. Variables’ significance and importance were consistent with the results for 30-day mortality. Age, sedated/paralysed/Glasgow coma scale (GCS) and source of admission/urgency of surgery were the strongest predictors. However, severe conditions in the medical history (such as metastatic disease and severe liver disease) and BMI had a more relevant role in the model (see Appendix 2, Table 41). The model showed good discrimination in the validation data set, a c index of 0.883 (95% CI 0.880 to 0.884) a Brier’s score of 0.1045 and a SMR of 1.00 (95% CI 0.99 to 1.01), which indicated good calibration (see Table 12 and Figure 5). The final coefficients are shown in Appendix 3, Table 46.

Mortality at 1 year following critical care admission

Between January 2013 and December 2013, there were 144,720 first admissions to 235 adult critical care units in England participating in the CMP with identifiable linkage with HES and death registrations. After the exclusions, a total of 127,855 were included in the development data set.

In total, 38,191 (29.9%) patients died during the one-year follow-up. Of these, 18,038 (14.1%) died during the critical care admission.

Baseline severe and chronic conditions of the included patients and associated 1-year mortality are described in Table 14. The most common severe condition in the medical history was immunocompromise (6%) and the highest 1-year mortality was associated with haematological malignancy (62%). Regarding RCS Charlson comorbidities, chronic pulmonary disease (14%), any malignancy (13%) and diabetes mellitus (12%) were the most prevalent. Dementia (48%), metastatic solid tumour (48%) and chronic renal disease (44%) were associated with the highest 1-year mortality. Figure 8 shows the Kaplan–Meier survival curve for 1-year survival after critical care admission.

FIGURE 8.

Kaplan–Meier survival curve for 1 year following critical care admission.

Model development

Functional form and significance of CMP physiological and non-physiological predictors previously included in the existing model for acute hospital mortality were reassessed. Consistency with global significance of the existing predictors was found after fitting a model considering all physiological and non-physiological variables from the current ICNARC model. The functional form of the continuous physiological predictors was reassessed for the 1-year mortality outcome. In all cases, predictors showed significant non-linearity (p < 0.001). The optimal functional form selected to model the continuous physiological predictors was four knots for heart rate, systolic blood pressure and creatinine. For temperature, urine output, PaO₂/FiO₂, urea, white blood cell count, potassium, glucose, blood lactate, platelet count, neutrophil count and urine output extending to five knots was needed. For arterial pH, a simplification using three knots was enough to accommodate the non-linear behaviour. Finally, respiratory rate and PaCO₂ were modelled using right-restricted cubic splines. This was necessary to capture the initial decrease in mortality and ‘spoon’ behaviour. This approach had a better fit than five knots and was more plausible than four.

Age and BMI were modelled as continuous, non-linear relationships using restricted cubic splines with five knots.

Following the development of the main model, severe conditions in the past medical history (APACHE II) and RCS Charlson comorbidities were added into the model. As key predictors for long-term mortality, the significance and importance of the comorbidities are shown in Appendix 2, Table 42. All APACHE II severe conditions and most RCS Charlson comorbidities were retained because of their significant effect and contribution to the model, except previous myocardial infarction (MI), dementia, hemiplegia or paraplegia, diabetes mellitus and rheumatological disease. The incorporation of APACHE II and RCS Charlson comorbidities, in particular metastatic disease and severe liver disease, in the risk prediction model produced better fit (Table 15). The comorbidities with the greatest magnitude of association were metastatic disease, severe liver disease and hematological malignancy (Table 16).

TABLE 15 - Model performance for predicting mortality at 1 year following critical care admission and for predicting mortality at 1 year following hospital discharge in the development cohort

Physiological and non-physiological predictors from the CMP previously included in the existing model to predict mortality at discharge from acute hospital.

Final risk model for mortality at one year following critical care admission excluding comorbidities.

Model	df	LL	BIC	c index (95% CI)	Brier’s score
One year following critical care admission
Main modela	81	–45347.44	91637	0.825 (0.822 to 0.828)	0.149
Main model + comorbidities	100	–43853.09	88867	0.839 (0.836 to 0.842)	0.144
Main model + comorbidities + reason for admission	167	–42429.44	86790	0.851 (0.848 to 0.854)	0.139
Main model + comorbidities + reason for admission + interactions	292	–41992.53	84571	0.854 (0.851 to 0.856)	0.138
One year following hospital discharge
Main modelb	81	–26226.97	53376	0.721 (0.716 to 0.726)	0.102
Main model + comorbidities	100	–24908.08	50952	0.766 (0.761 to 0.771)	0.097

TABLE 16 - Odds ratios for comorbidities in the risk model for mortality at 1 year following critical care admission and in the risk model for mortality at 1 year following hospital discharge

OR, odds ratio.

Comorbidity	OR (95% CI) for 1-year mortality following critical care admission	OR (95% CI) for 1-year mortality following hospital discharge
Previous MI	1.12 (1.00 to 1.23)	1.14 (1.00 to 1.28)
Congestive cardiac failure	1.28 (1.19 to 1.38)	1.40 (1.27 to 1.54)
Peripheral vascular disease	1.19 (1.10 to 1.28)	1.23 (1.12 to 1.35)
Very severe cardiovascular disease	1.30 (1.15 to 1.46)	1.17 (0.98 to 1.37)
Cerebrovascular disease	1.16 (1.05 to 1.28)	1.17 (1.00 to 1.33)
Dementia	1.13 (0.97 to 1.32)	1.37 (1.13 to 1.65)
Hemiplegia or paraplegia	1.07 (0.88 to 1.28)	1.22 (0.95 to 1.55)
Chronic pulmonary disease	1.13 (1.07 to 1.19)	1.21 (1.13 to 1.29)
Severe respiratory disease/home ventilation	1.73 (1.55 to 1.92)	1.62 (1.40 to 1.87)
Liver disease	1.46 (1.31 to 1.63)	1.44 (1.24 to 1.66)
Severe liver disease	2.39 (2.16 to 2.63)	1.87 (1.61 to 2.17)
Chronic renal disease	1.25 (1.15 to 1.34)	1.29 (1.17 to 1.42)
End-stage renal failure	1.73 (1.55 to 1.92)	1.56 (1.30 to 1.87)
Any malignancy	1.60 (1.51 to 1.69)	1.89 (1.76 to 2.02)
Haematological malignancy	2.15 (1.89 to 2.42)	2.11 (1.79 to 2.47)
Metastatic disease	4.04 (3.74 to 4.36)	4.46 (4.10 to 4.85)
Immunocompromise	1.47 (1.36 to 1.57)	1.58 (1.45 to 1.71)
Diabetes mellitus	1.03 (0.97 to 1.08)	1.17 (1.09 to 1.25)
Rheumatological disease	1.01 (0.90 to 1.12)	1.01 (0.87 to 1.16)

After adjusting for comorbidities, deprivation was not retained in the main term model because of a low contribution to the model fit.

A total of 56 process/system combinations and 18 individual conditions from the ICNARC Coding Method (Table 17) were selected after the modelling process described above. Comparing with the ICNARC acute hospital mortality and the 30-days primary reason for admission categories, the selected individual conditions were less acute and more cancer-related. Significant interactions with physiology (n = 24) were incorporated to the model.

TABLE 17 - Comparison of individual reasons for admission included in the risk models for acute hospital mortality and 1-year mortality

Shaded cells signify reasons for admission that were included in the risk model for 1-year mortality and not in the risk model for acute hospital mortality.

Acute hospital mortality	One-year mortality
Toxic or drug-induced coma or encephalopathy	Toxic or drug-induced coma or encephalopathy
Lower limb artery stenosis or occlusion	Lower limb artery stenosis or occlusion
Anaphylaxis	Anaphylaxis
Leaking large bowel anastomosis	Leaking large bowel anastomosis
Acute alcoholic hepatitis/alcoholic cirrhosis	Acute alcoholic hepatitis/alcoholic cirrhosis
Thrombo-occlusive disease of brain	Thrombo-occlusive disease of brain
Secondary hydrocephalus	Secondary hydrocephalus
Pulmonary fibrosis or fibrosing alveoli	Pulmonary fibrosis or fibrosing alveolitis
Asthma attack in new or known asthmatic	Asthma attack in new or known asthmatic
Hanging or strangulation	Hanging or strangulation
Anoxic or ischaemic coma or encephalopathy	Anoxic or ischaemic coma or encephalopathy
Entero-enteric or entero-cutaneous fistula	CABG for chronic angina
Fractured ribs	Pancreatic or pancreato-duodenal tumour
Fungal or yeast pneumonia	Secondary hepatic tumour
Haemolysis or thrombocytopaenia	Small bowel tumour
Intracerebral haemorrhage	Malignant large bowel tumour
	Large bowel tumour
	Carotid or vertebral artery stenosis or occlusion

Interactions included in the previous model between severe liver disease with physiology, CPR with physiology, ventilation with physiology and physiology with physiology appeared to be less important for 1-year mortality and only the interactions between CPR and temperature and between ventilation and respiratory rate, PaO₂/FiO₂ and PaCO₂ were included after adjusting for the main term model plus the new primary reason categories and interactions.

The final model showed good performance (c index of 0.855 and Brier’s score of 0.118) and internally, calibration of the model was satisfactory (Figure 9). Overfitting was of limited relevance because of the very large data set and, as expected, model optimism was negligible (0.77% estimated overfitting). Full coefficients for the final model are presented in Appendix 3, Table 47.

FIGURE 9.

Calibration of the final risk model for mortality at one year following critical care admission in (a) the development data set and (b) the validation data set.

Mortality at 1 year following hospital discharge

A total of 100,450 admissions discharged alive from hospital between January 2013 and December 2013 were selected for the hospital survival cohort. With the aim of evaluating if the effect of comorbidities is different among hospital survivors, their significance and contribution were retested in a model considering 1-year survival from hospital discharge as outcome and adjusting by the variables resulting in the previous main term model.

Comorbidities were present among hospital survivors in similar proportions to the critical care admission cohort (see Table 14). In general, comorbidity remained in importance as predictors of 1-year mortality (see Appendix 2, Table 42) and showed a similar effect (see Table 16), with dementia and diabetes mellitus arising as significant and important predictors. On the other hand, and liver disease and haematological malignancy became less important predictors of 1-year mortality for hospital survivors. Very severe cardiovascular disease and cerebrovascular disease were not significant after adjusting for main terms.

The c index of the main terms model without and with comorbidities were 0.721 (95% CI 0.716 to 0.726) and 0.766 (95% CI 0.761 to 0.771), respectively. The addition of comorbidities showed a greater improvement of the performance of the model (see Table 15) in the hospital survivor cohort, although the performance was better when predicting 1-year mortality in the critical care admission cohort.

Study cohort

We selected NHS adult critical care units in England participating in the CMP with identifiable linkage with HES and death registrations. False linkage and errors in linkage with HES and death registrations were excluded. The linked cohort included patients discharged from hospital between 1 April 2009 and 31 March 2016 following a critical care admission. The final follow-up date was 31 December 2016 (based on the latest available data from the UKRR at the time of linkage). For patients with multiple hospital episodes that included critical care admissions, we considered the index admission to be the first hospital admission during the analysis period. Re-admissions to critical care and transfers during the index admission were excluded.

Inclusion and exclusion criteria

Patients included in the model were all those discharged alive from acute hospital following a critical care unit admission. Patients were excluded if they had pre-existing ESRD, identified by:

• linkage with the UKRR indicating a date of diagnosis of ESRD prior to the date of discharge from hospital of the index admission

• ICD-10 codes indicating ESRD (I120, I129, N186, Z49, Z940, Z992) in any diagnosis field from linked HES records from a hospital episode either prior to or during the index admission

• recording of an ongoing requirement for RRT for irreversible ESRD in the CMP.

Outcome

The main outcome was new receipt of RRT for ESRD, based on the date of diagnosis recorded in the UKRR database, after the date of discharge from hospital. The competing risk was death from any cause before ESRD, identified via linkage with death registrations. Deaths after ESRD were not considered.

Candidate variables

As the important risk factors and their relationships with the outcome may be very different from those considered for other outcome measures, risk models for ESRD were developed de novo, using the same methods as previously applied to develop the original risk models for acute hospital mortality in adult critical care. Potentially important candidate predictors were chosen based on expert clinical opinion and availability in the linked data sources. RCS Charlson comorbidities were identified as described in Chapter 3, except for CKD and AKI, for which the lookback period was extended to 5 years. A description of the candidate predictors is given in Appendix 1, Table 38.

Statistical analyses

For the description of cohort characteristics, assessment of optimal functional form and selection of covariables, we followed the same methodology as described in Chapter 3.

The incidence rate was calculated as the number of new cases of ESRD divided by the follow-up time and expressed as the number of events per 1000 person-years. Because outcomes of death and ESRD have an important competing effect, analysis of the predictors of ESRD needs to consider the risk of dying before ESRD. In the competing risk context, two different approaches to regression modelling exist: the Cox proportional cause-specific regression model and the Fine-Gray regression model. The former models the dependence of the cause-specific hazard function on covariates and the latter models the dependence of the cumulative incidence function (CIF) on predictors. Both methods were considered pertinent in the present study as we aimed to explore the role of acute and chronic risk factors on development of ESRD following a critical care admission. For the cause-specific model, the explained variation (R²) was determined. The proportional hazard assumption was tested by visual inspection of Schoenfeld residual plots and log-log plots.

Patient characteristics

Characteristics of the overall cohort, as well as divided by outcome, are detailed in Table 18. Compared with the overall cohort, patients who developed ESRD were similar in age (mean 60 years), with a higher proportion of males (62% vs. 55%) and of non-white ethnicities (16% vs. 7%). More than 30% suffered from CKD in either the current hospital admission or an admission during the previous 5 years (compared with 3.6% for the overall cohort) and 6% (vs. < 1%) had previous hospital admissions involving AKI. The most prevalent non-renal chronic condition in patients who developed ESRD was diabetes (24% vs. 10%). Patients who developed ESRD had higher creatinine levels during the first 24 hours following admission to critical care (mean 530 vs. 111 µmol l⁻¹) and urea (mean 27 vs. 8 µmol l⁻¹) and lower urine output (mean 1354 vs. 2013 ml) and more than half had received renal support during the critical care unit stay (51% vs 5%).

Predictors of end-stage renal disease

Table 19 shows the cHR after the modelling process. Although pregnancy was not associated with development of ESRD, it was kept for confounding. An increasing heart rate, respiratory rate, arterial pH, white blood cell count, PaCO₂ and urine output during the first 24 hours of critical care were all linearly associated with a decrease in the risk of developing ESRD. On the other hand, increasing systolic blood pressure was significantly associated with an increased risk of ESRD. Non-linear associations were observed between the risk of ESRD and the following factors: age, sodium, creatinine, urea and blood lactate. These factors were modelled by restricted cubic splines. Figure 12 shows the adjusted cause-specific graphs associated with these non-linear relationships to facilitate the presentation and interpretation. There was a huge increase in the risk of ESRD when creatinine (measured during the first 24 hours of critical care) increased over a range from 100 to 200 µmol l⁻¹. Patients with CKD [adjusted hazard ratio (aHR) 4.11], AKI (aHR 1.73), peripheral vascular disease (aHR 1.29) and diabetes (aHR 1.35) were more likely to receive RRT for ESRD than those without these conditions. Mechanical ventilation (aHR 1.29), vascular surgery (aHR 1.23) and nephrectomy (aHR 1.92) were also associated with an increased risk of ESRD. However, some severe conditions in the medical history [severe liver disease (aHR 0.81) and metastatic disease (aHR 0.65)] were associated with a lower likelihood of RRT for ESRD. The effects of the rest of the variables in the model are presented in Table 19. CPR prior to admission, cardiac surgery, severe cardiovascular disease, MI and congestive cardiac failure were not considered significant after adjusting for the rest of the variables in the model. No significant interactions were demonstrated.

FIGURE 12.

Cause-specific hazard ratio for (a) ESRD and (b) mortality before ESRD after hospital discharge following critical care for continuous predictors included in the model as non-linear using restricted cubic splines. ^a a, Physiological parameters assessed during the first 24 hours following admission to critical care.

Harrell’s c-statistic for the model was 0.94, with an explained variation (R²) of 0.985 (95% CI 0.982 to 0.987, based on 100 bootstrap samples). Most of the prognostic information was carried by duration of renal support, surgical status, CKD and highest creatinine, because the R² decreased considerably on dropping those variables (see Report Supplementary Material 1, Table S1).

Predictors of cumulative incidence of end-stage renal disease

Results of the competing risk analysis using the Fine–Gray approach were consistent with the cause-specific Cox results. Significant predictors of cumulative incidence of ESRD and the corresponding adjusted sHR are presented in Table 19. Figure 13 provides a graphical interpretation of the non-linear association between continuous factors and the cumulative incidence (absolute risk over time) of the development of ESRD. Cumulative incidence of ESRD increases with increasing age and urea, particularly in admissions with elevated creatinine. On the other hand, admissions with higher blood lactate had a lower risk of RRT for ESRD (adjusted for other variables in the model). The presence of severe liver disease (sHR 0.70) and metastatic disease (sHR 0.41) were significantly associated with a lower risk of RRT for ESRD. CKD (sHR 3.58) was associated with a substantial increase in the cumulative incidence of ESRD (Figure 14). Previous AKI (sHR 1.64), diabetes (sHR 1.30) and peripheral vascular disease (sHR 1.25) were also all associated with a higher risk of ESRD.

FIGURE 13.

Cumulative incidence of ESRD in the 7 years after hospital discharge following critical care according to predictors included in the Fine–Gray competing risks model. (a) Age; (b) highest creatinine; (c) highest sodium; (d) highest urea; (e) highest blood lactate. ^a a, Physiological parameters assessed during the first 24 hours following admission to critical care.

FIGURE 14.

Cumulative incidence of ESRD in the 7 years after hospital discharge following critical care according to predictors included in the Fine–Gray competing risks model (comorbidities). (a) CKD; (b) AKI; (c) PVD; (d) diabetes; (e) severe liver disease; (f) metastatic. PVD, peripheral vascular disease.

Study cohort

The cohort for this chapter is the CMP hospital survivor cohort (see Chapter 3) of patients discharged from hospital between 1 April 2009 and 31 March 2016 following a critical care episode. The final follow-up date was 31 March 2017 (based on the latest available data from the NDA at the time of linkage).

Exclusion criteria

Patients were excluded if they had pre-existing diabetes (type 1 or type 2), identified by:

1. linkage with the NDA indicating a date of diagnosis prior to or during the critical care episode

2. ICD-10 codes indicating diabetes in any diagnosis field from linked HES records prior to or during the critical care episode

3. a primary or secondary reason for admission to the critical care unit in the CMP associated with diabetes.

Outcome

The main outcome was incidence of type 2 diabetes. As the actual timing of the development of diabetes is unknown, we used as a surrogate measure a new registration for type 2 diabetes, based on the date of diagnosis recorded in the NDA database, after the date of discharge from hospital. The competing risk was death from any cause before type 2 diabetes, identified via linkage with death registrations. Deaths after a diagnosis of type 2 diabetes were not considered. The association between glucose and mortality was also explored as a secondary outcome.

Candidate variables

Potentially important candidate predictors and controlling variables were chosen based on expert clinical opinion and availability in the linked data sources. A description of the candidate predictors is given in Appendix 1, Table 38. The primary aim was to evaluate the association between serum glucose levels in the first 24 hours after admission and subsequent risk of type 2 diabetes.

Statistical analyses

For the description of cohort characteristics, assessment of optimal functional form and selection of covariables, we followed same methodology as described in Chapter 3.

The incidence rate was calculated as the number of cases of new-onset type 2 diabetes divided by the follow-up time and expressed as the number of events per 1000 person-years. The hazard ratios and cumulative mortality function were assessed by Cox (cause-specific hazard) and Fine–Gray (subdistribution hazard) methods; both approaches were described in Chapter 6. For the cause-specific model, the explained variation (R²) was determined. The proportional hazard assumption was tested by visual inspection of Schoenfeld residual plots and log-log plots.

Patient characteristics

Characteristics of the overall cohort, as well as divided by outcome, are detailed in Table 20. Compared with the overall cohort, patients with subsequent type 2 diabetes were slightly older (mean age 62 vs. 59 years), with a higher proportion of males (60% vs. 55%) and higher BMI (mean 30.7 vs. 26.7 kg/m²). Five per cent of the patients who developed type 2 diabetes were Asian compared with only 3% in the overall cohort. There was also a much stronger deprivation gradient across those who developed type 2 diabetes (from 15% in the least deprived quintile to 26% in the most deprived quintile) compared with the overall cohort (18% to 22%). Severity scores were similar for patients who developed type 2 diabetes and the overall cohort, but there were differences in some individual physiological parameters. In particular, the patients who went on to develop type 2 diabetes had higher serum glucose levels (mean 10.2 vs 8.5 mmol l⁻¹) during the first 24 hours in the critical care unit.

Data and resource use

The patients included in the prediction model were those discharged alive from hospital between January 2013 and December 2014. The resource use was measured as the number of days in acute hospital care during subsequent hospital episodes (as identified through data linkage with HES) from the hospital discharge until 1-year follow-up or death within that period. For hospitalisations in which the patient was admitted and discharged on the same day, the number of days in hospital was 0. For those subsequent hospitalisations in which the patient was discharged more than 1-year after index discharge, only the days in hospital occurring during the study period were considered. Total subsequent cost was calculated by summing the cost for subsequent hospitalisations and the cost for subsequent critical care admissions. The cost for each subsequent hospitalisation was measured as the cost of the full hospital spell.

The cost of the full hospital spell was valued using the Department of Health and Social Care’s (DHSC) APC tariff. Information from the HES episodes files were used in the cost for hospitalisation calculation. The dominant health-care resource group (HRG) was identified from HES data [Secondary Uses Service (SUS)-generated core spell HRG]. When the HES field was not available, the HRG attributed to the admission episode was used. The bed-day cost for each HRG was assigned using the DHSC APC tariff. 72 Total spell length of stay was identified from the spell’s discharge information (spell duration) in the final episode of the spell. When the patient was discharged more than 1 year after the index discharge, only the days in hospital occurring during study period were considered. Days in critical care were removed from the overall spell duration as critical care has a different tariff. Each HRG was matched to their corresponding trim point for determining long versus short stays for each spell. We used patient classification and the admission method from the HES data set to identify day case and elective inpatient stay. Those HRGs with a locally determined cost were identified and they were not included in this calculation. We assumed that the reference costs for the same period were proxies for the tariff for non-attached HRGs. Hence, these were cross-matched from the HRG4 to the HRG4+ currency design and valued using spell-level NHS reference costs. 15 Finally, missing or unlinked HRGs were valued using the average by type of admission of all HRGs.

Statistical analysis

Days in hospital and cost were reported as described in Chapter 3.

To assess the predictors of costs during the year after the index hospital discharge, a two-part regression model was used. Health-care costs usually have distributions that are skewed with a large mass at zero. In modelling such outcomes, a two-part model has become a best-practice approach, improving the fit of the model and allowing for better understanding of results. 73 A logistic regression was first used to assess the predictors of having health-care cost (at least one hospital and/or critical care admission) during the year following index hospital discharge. Then, conditional on having any health-care cost, a generalised linear model (GLM) with a gamma distribution and a log-link function was selected to determine the predictors of 1-year cost after index hospital discharge. We checked the log link against several other functional form alternatives. In addition, a modified ‘Park Test’74 was used to test the distribution family. Non-linearity, inclusion of variables and interactions were assessed using methods described in Chapter 3. As recommended,75 we applied the general practice that any variable that is in either the first part or the second part model will be in both. No variables are included in one part but excluded from the other. In addition, final selected variables were tested for jointly significant in both parts of the two-part model. Predicted values and marginal effects were then calculated accounting for the full model.

Baseline characteristics

The baseline characteristics of the included patients are described in Table 22. The median age was 64 years (IQR 49–75 years) and more than a half of patients were male. Most of the patients were able to live without assistance in daily activities (79.4%). The median critical care length of stay of the cohort was 48 hours (IQR 23–103 hours) and the median hospital length of stay was 12 days (IQR 7–25 days).

Subsequent hospital/critical care admission during the first year and estimated health-care cost

Results are summarised in Table 23. The rate of subsequent hospitalisations for the first year was 1.06 hospitalisations per patient. The mean health-care cost during the first year after index hospital discharge was £3734, with over half of costs having a value of zero. The distribution of total cost was highly skewed with a large mass at zero (Figure 21). A total of 97,593 patients (47%) had a subsequent health-care cost (hospitalisation/critical care admission) with a mean cost of £7952 (median £4566, IQR £2288–9587). When we split by source of cost, both hospitalisation and critical care costs had a large proportion of zeros and a declining distribution of positive costs (see Figure 21). For the 97,229 (47%) individuals with a subsequent APC cost, the mean cost was £6589 (median £4204, IQR £2207–8482). A total of 14,293 (6.9%) patients were admitted to critical care during the first year after index hospital discharge, with a mean cost of £9465 (median £4142, IQR £2761–9436).

FIGURE 21.

Distribution of total health-care cost: full distribution including zeros, the histogram of just positive values and split by APC and critical care cost. (a) Total health-care cost; (b) total health-care cost if > £0; (c) total APC health care cost if > £0; (d) total critical care health-care cost, if > £0.

Factors associated with non-zero health-care cost during the first year

After testing for linearity, the continuous variables age, BMI and ICNARC physiology score were modelling using restricted cubic splines. The specification test supported the use of the log-link and the gamma distribution. Results of the two-part model are presented in Appendix 3, Table 48. The predictors of whether the patient would have any subsequent health-care costs were previous hospitalisation, critical care length of stay, age, BMI, illness severity, mechanical ventilation, dependency prior to admission, source of admission, CPR and deprivation. Regarding severe conditions in the medical history, severe liver disease, metastatic disease, haematological malignancy, severe respiratory disease and ESRD were also found to be significant predictors. The model indicated that people with the following comorbidities were also more likely to have non-zero health-care costs in the year following hospital discharge: previous MI, congestive cardiac failure, peripheral vascular disease, dementia, chronic pulmonary disease, rheumatological disease, liver disease, diabetes mellitus, hemiplegia or paraplegia, chronic renal disease and any malignancy.

Factors associated with health-care cost during the first year, conditional on having non-zero cost

Conditional on hospital/critical care admission, the GLM identified similar predictors for health-care cost as those for non-zero cost, apart from CPR, previous MI and dementia. Patients who had previous hospitalisations and patients who required total assistance with daily activities had increased costs during the first year. In general, having severe conditions in the medical history or chronic conditions were found to incur significantly higher costs during the first year, but this declined with age.

Table 24 shows the combined marginal effects from both parts of the model. The predicted mean total cost was £3725 per person in the year following index hospital discharge, which is close to the observed mean of £3708. Previous hospitalisation was associated with an increase in health-care costs of £999. Patients with some or total assistance with daily activities also had higher costs. The results show that all severe or chronic conditions are associated with greater costs. To better understand the effect of the continuous variables and interactions, we plotted the marginal effect of the continuous predictors on the health-care costs (Figures 22–24).

FIGURE 22.

Marginal plot of the effect of BMI on total health-care costs.

FIGURE 23.

Marginal plot of the effect of ICNARC physiology score and interactions with comorbidities on total health-care cost. (a) Overall marginal effect; (b) severe liver disease in medical history; (c) metastatic disease; (d) chronic pulmonary disease. Solid line: presence of comorbidity. Dashed line signifies absence of comorbidity.

FIGURE 24.

Marginal plot of the effect of age and interactions with comorbidities on total health-care cost (solid line: presence of comorbidity). (a) Overall marginal effect; (b) any malignancy; (c) haematological malignancy; (d) metastatic disease; (e) immunocompromise; (f) chronic pulmonary disease; (g) end-stage renal failure; (h) diabetes mellitus; (i) severe respiratory disease; (j) liver disease. Dashed line signifies absence of comorbidity.

Study cohort

For the development data set, we selected NHS cardiothoracic critical care units in England participating in the CMP, with identifiable linkage with HES and death registrations between 1 April 2009 and 31 March 2015. The validation data set consisted of admissions to cardiothoracic critical care units between 1 April 2015 and 31 March 2016.

Inclusion and exclusion criteria

Patients included in the model were all admissions following cardiothoracic surgery identified from data linkage with NACSA, excluding any with a date of admissions preceding the NACSA procedure date or not in the first 20 days after surgery. Patients aged < 16 years, re-admissions to the critical care unit and patients transferred from another critical care unit were excluded.

Outcome

Risk models were developed for two outcomes: acute hospital mortality and 1-year mortality. Acute hospital mortality was defined as death before final discharge from acute hospital and included deaths after direct transfer to another acute hospital from the hospital housing the critical care unit. One-year mortality was obtained by data linkage with death registrations.

Candidate predictors

Candidate predictors were chosen based on the previously developed risk prediction model for acute hospital mortality among admissions to cardiothoracic critical care units,16 pre- and intra-operative risk factors obtained by data linkage with NACSA, and APACHE II and RCS Charlson comorbidities (see Appendix 1, Table 39). Severe conditions in the past medical history (APACHE II) and RCS Charlson comorbidities were identified and defined as described in Chapter 3.

Development of risk model for acute hospital mortality

Between 1 April 2009 and 31 March 2016, there were 40,516 admissions to seven cardiothoracic critical care units in England participating in the CMP with identifiable linkage with HES and death registrations. Of these, 31,415 (78%) were successfully linked with NACSA. After excluding admissions not following cardiac surgery, a total of 27,687 eligible admissions to seven cardiothoracic critical care units formed the development data set (Figure 25).

FIGURE 25.

Flow diagram.

A total of 1072 (3.9%) admissions to cardiothoracic units died during the hospitalisation. Of these, 693 (2.5%) died during the critical care unit admission. The median hospital length of stay of the cohort was 10 days (IQR 7–16 days), with a median critical care unit length of stay 26 hours (IQR 21–68 hours). Cohort characteristics are shown in Table 25.

During our evaluation, the following variables were considered redundant or inoperative because of either high proportions of missing values or small sample sizes in the category of interest and were not included in the analysis: ventricular assist device used (pre operative), aortic valve procedure, mitral valve procedure, tricuspid valve procedure, pulmonary valve procedure, intra-aortic balloon pump used (intra operative), impeller device used (intra operative), ventricular assist device used (intra operative), other support device used (intra operative), aortic pathology – root segment, aortic pathology – ascending segment, aortic pathology – arch segment, aortic pathology – descending aorta segment, aortic pathology – abdominal segment, systolic pulmonary artery pressure (PA systolic), total number of distal coronary anastomoses and number of valves replaced/repaired.

A multivariable model with the factors included in the previous risk prediction model for acute hospital mortality among admissions to cardiothoracic critical care units was refitted, excluding location and cardiac surgery. The significance and functional forms for the remaining predictors were retested. The completeness for those variables was 98.2% and consequently data were not imputed. All predictors were significant, and their functional forms were consistent with the previous findings. Sex was then including and the c index and Brier’s score for the main model were 0.873 and 0.0280, respectively. Comorbidities, pre-operative and factors from NACSA that were significant (p < 0.1) in univariable analyses were introduced into the main model and a final model was developed as described above.

The following factors from NACSA were found to be associated with acute hospital mortality on multivariable analysis: diabetes (categorised as no diabetes vs. diabetes controlled by diet/oral therapy/insulin), atrial fibrillation/flutter, dyspnoea status pre surgery (categorised as no limitation or slight limitation of ordinary physical activity vs. marked limitation of ordinary physical activity vs. symptoms at rest or minimal activity), history of pulmonary disease, history of neurological dysfunction; extracardiac arteriopathy, operative urgency (categorised as elective vs. urgent vs. emergency vs. salvage), and cumulative bypass time. Of the comorbidities and pre-existing conditions, only severe respiratory disease, severe cardiovascular disease and congestive heart failure were significant predictors of acute hospital mortality.

The first expanded model, incorporating factors from NACSA, performed moderately better than the baseline parsimonious model (Table 26). The second expanded model with severe conditions and pre-existing congestive heart failure resulted in minimal change to c index and Brier’s score, with a c index of 0.892 and Brier’s score of 0.0273 for the final model. Appendix 2, Table 43, summarises the significance and importance of the predictors in the final model. Full coefficients for the final model are presented in Appendix 3, Table 49. After comparing the reclassification of the two models using risk categories defined by thresholds of 0%, 2%, 5%, 10%, 20% and 50% (Tables 27 and 28), a total of 5766 (21%) admissions were reclassified and 3088 of those (54%) were placed in more appropriate categories. The total NRI for the expanded model was 7.0% (standard error 2.0%, p < 0.0005). The distribution of predicted acute hospital mortality from the new model is shown in Figure 26.

FIGURE 26.

Distribution of predicted risk of acute hospital mortality.

Development of risk model for 1-year mortality

For the development of the 1-year mortality model we used the same cohort that we used for the acute hospital mortality. Two admissions were excluded owing to missing date of death. The median follow-up was 5 years (IQR 3.9–6.4 years). In total, 1918 (6.92%) died during the 1-year follow-up, but most of the deaths were in the first month from admission. A Kaplan–Meier plot of time to death within 1 year is shown in Figure 30.

FIGURE 30.

Kaplan–Meier plot of time to death within 1 year from admission to cardiothoracic unit.

The risk model developed for in hospital mortality was refitted to the new outcome and all predictors were significantly associated. The model showed a good performance, with a c index of 0.824 (95% CI 0.814 to 0.834) and a Brier score of 0.0516. After the addition and reassessment of the predictors that were found not to be important predictors for acute hospital mortality, the following pre-operative factors from NACSA were included in the final model: renal function/dialysis [recategorised as none or dialysis for acute renal failure vs. dialysis for chronic renal failure vs. no dialysis but pre-operative acute renal failure (anuria or oliguria < 10 ml/hour)]; left ventricular ejection fraction, number of previous MIs and major aortic procedure. No additional pre-existing conditions were found to be important predictors. After accounting for these predictors, the discriminative ability of the model improved slightly with a c index of 0.827 (95% CI 0.816 to 0.837). The full model coefficients are presented in Appendix 3, Table 49.

Following the development process described above, the significance and importance of the predictors in the final model are shown in Appendix 2, Table 43. Age was one of the strongest predictors of 1-year mortality, followed by GCS, blood lactate and operative urgency. Compared with the short-term model (acute hospital mortality), pre-operative and pre-existing conditions had a more important role in the model performance, whereas physiological and laboratory variables were stronger predictors of acute hospital mortality.

Methods

Methods common to all objectives and analyses were describe in Chapter 3. Details relating to the study cohorts and outcomes can be found in the same chapter.

Results

One hundred and eighty-one hospitals participated in NCAA between 1 January 2013 and 31 December 2014. During this time there was a total of 33,829 team visits following 2222 calls for cardiac arrest with identifiable linkage with HES and death registrations. After removing records that were ineligible for inclusion in the prediction models, there were a total of 26,904 eligible patients and, of these, 26,748 (99.4%) patients from 172 hospitals were included in the modelling. The breakdown of the inclusion/exclusion process for the development data set is shown in Figure 32. Characteristics and outcomes of in-hospital cardiac arrest patients in the development and validation data sets are summarised in Table 30.

FIGURE 32.

Flow diagram for cohort identification.

Validation of prediction models for return of spontaneous circulation > 20 minutes and hospital survival

After exclusions, a total of 7073 patients experiencing an in-hospital cardiac arrest between 1 January 2015 and 30 June 2015 comprised the validation cohort. As for the development cohort, discrimination and accuracy were better for hospital survival (c index 0.82, Brier’s score 0.116) than for ROSC > 20 minutes (c index 0.72, Brier’s score 0.212). Calibration plots showed good calibration for the validation cohort (Figure 33; see Report Supplementary Material 1, Figure S4).

FIGURE 33.

Calibration plot for ROSC > 20 minutes, hospital survival and for 1-year survival (validation set). (a) Hospital survival; (b) ROSC > 20 minutes; (c) 1-year survival.

Development and validation of a prediction model for 1-year survival

A total of 4454 (16.7%) patients were alive after 1 year following an in-hospital cardiac arrest (Figure 34), 84% of 60-day survivors. All predictors, excluding pre-existing comorbidities, were reassessed for 1-year survival and were found to be significant (see Report Supplementary Material 1, Table S2). After their addition into the model, the following pre-existing comorbidities that were found significant for hospital survival significantly predicted 1-year survival following an in-hospital cardiac arrest: congestive cardiac failure, peripheral vascular disease, severe liver disease, malignancy and metastatic solid tumour. Chronic renal disease and dementia were also found to be significant predictors of 1-year survival, but hemiplegia or paraplegia was not (see Appendix 2, Table 44). Metastatic solid tumour, malignancy, congestive cardiac failure, severe liver disease and chronic renal disease had a relevant role in predicting 1-year survival. The 1-year survival model coefficients are presented in Appendix 3, Table 52. The model c index after addition of comorbidities was 0.829 (95% CI 0.822 to 0.835). The model showed good discrimination in the validation cohort, with a c index of 0.816 (95% CI 0.799 to 0.833) and a Brier score of 0.116 (see Table 32), and acceptable calibration (see Figure 33 and Report Supplementary Material 1, Figure S4).

FIGURE 34.

Kaplan–Meier survival estimate to 365 days following start cardiac arrest.

Methods

Methods common to all objectives and analyses were describe in Chapter 3. Details relating to the study cohorts and outcomes can be found in the same chapter.

Results

There were 64,871 validated team visits between 1 April 2011 and 31 March 2015. After removing cardiac arrest non-survivors, second and subsequent team visits to the same patient, patients < 16 years and arrests in critical care, 18,410 first team visits (patients) were considered for linkage. A total of 4864 were successfully linked with CMP, which corresponded to 5594 unique critical care unit admissions, including transfers and re-admissions. The outcome of total critical care unit length of stay during the same hospital stay was then calculated for each patient. Only the information from the first critical care unit admission was then kept in the analysis. After excluding 23 patients missing all physiology, a total of 4841 patients were entered into the analysis.

Characteristics and outcomes of survivors of in-hospital cardiac arrest admitted to critical care are summarised in Table 35, split by hospital survivors and non-survivors. A total of 1839 (38.0%) patients admitted to critical care following in-hospital cardiac arrest were alive at hospital discharge and 3002 (62.0%) were hospital non-survivors. The mean total critical care unit length of stay was 8 days for survivors and 4 days for non-survivors, with mean costs of approximately £13,000 and £7000, respectively.

Determinants of total critical care unit length of stay following in-hospital cardiac arrest in hospital survivors

Results of the multivariable analysis for total critical care unit length of stay following an in-hospital cardiac arrest in hospital survivors are presented in Appendix 3, Table 53. The following factors were significant in determining total critical care unit length of stay: location of arrest, presenting rhythm, reason for admission to critical care by body system, number of advanced organs supports received and severe conditions in the medical history. In addition, severity of illness, measured by the ICNARC Physiology Score, and age showed significantly non-linear associations with the total critical care unit length of stay. The interaction between severe conditions in the medical history and severity of illness also significantly influenced the total critical care unit length of stay following an in-hospital cardiac arrest.

The overall predicted mean for hospital survivors was 8.5 days (95% CI 8.0 to 9.2 days), being very similar to the observed mean. Taking into account the rest of the patients’ characteristics in the model, the mean for a patient that had an arrest on the ward would be 10 days but if they were located in theatre the mean is 6 days (Table 36). Patients presenting following pulseless electrical activity or bradycardia are likely to have a longer critical care unit length of stay than those who present following shockable rhythms or asystole. Figures 35 and 36 show the predicted values for age and for the ICNARC Physiology Score according to presence or absence of severe conditions in the medical history. We observe that for older patients the length of stay is shorter, and although the length of stay increases for sicker patients, the association is less strong for patients with severe conditions in their medical history.

TABLE 36 - Predicted mean critical care unit length of stay (days) for hospital survivors by characteristics included in the model

CCU, coronary care unit; ED, emergency department

	Predicted mean	95% CI
Overall predicted mean	8.5	8.0 to 9.2
Location of arrest
Ward, obstetrics area, other intermediate care area, clinic or non-clinical area	10.4	9.3 to 11.3
ED or emergency admission unit	7.4	6.6 to 8.3
Theatre and recovery	5.9	4.7 to 7.0
Cardiac catheter lab or CCU	7.2	6.1 to 8.4
Imaging department or specialist treatment area	7.7	5.9 to 9.4
Reason for admission by body system
Respiratory	11.3	9.7 to 12.8
Cardiovascular	7.5	6.9 to 8.17
Gastrointestinal	10.6	8.2 to 13.0
Neurological (including eyes)	8.5	6.8 to 10.1
Others	7.6	6.1 to 9.1
Number of advanced organs supports
0	4.2	3.6 to 4.8
1	5.9	5.4 to 6.4
2	8.4	7.9 to 9.0
3	12.0	10.8 to 13.2
4	17.0	14.3 to 18.8
Presenting rhythm
Shockable	6.9	6.2 to 7.6
Non-shockable	9.6	8.8 to 10.3

FIGURE 35.

Predicted mean critical care unit length of stay for (a) hospital survivors and (b) hospital non-survivors by age. LOS, length of stay.

FIGURE 36.

Predicted mean critical care unit length of stay for hospital survivors by severity of illness and severe conditions in the past medical history. (a) No previous comorbidities; (b) previous comorbidities. ICU, intensive care unit; LOS, length of stay.

Determinants of total critical care unit length of stay following an in-hospital cardiac arrest in hospital non-survivors

For non-survivors, only the following variables significantly influenced the total critical care unit length of stay following in-hospital cardiac arrest: number of advanced organ supports received, age and the ICNARC physiology score, both modelled using restricted cubic splines. The interaction between the ICNARC physiology score and number of advanced organ supports was significant, meaning that the effect of illness severity varied across the number of organs supported (see Appendix 3, Table 54).

The overall predicted mean total critical care unit length of stay for non-survivors was 4.5 days (95% CI 4.0 to 4.7 days). The association with age was similar to that for survivors but with a lower resource use (see Figure 35). Predicted total critical care unit length of stay decreased for sicker patients but decreased differently according to the number of organs supported (Figure 37).

FIGURE 37.

Predicted mean critical care unit length of stay for hospital survivors by severity of illness and number of advanced organ supports.