A multicentre randomised controlled trial of Transfusion Indication Threshold Reduction on transfusion rates, morbidity and health-care resource use following cardiac surgery (TITRe2)

Changes to study design after commencement of the study

There were no major changes to the study design after commencement. Some changes were made to eligibility criteria and outcomes, which are discussed in Changes to study eligibility criteria after commencement of the study and Changes to study outcomes after commencement of the study, respectively.

Eligibility criteria

The study inclusion criteria were:

• adults of either sex, aged ≥ 16 years, undergoing cardiac surgery [defined as coronary artery bypass grafting (CABG), heart valve replacement or repair, aortic surgery or surgical correction of congenital cardiac disease]

• post-operative haemoglobin level < 9 g/dl at any stage during the patient’s post-operative hospital stay [i.e. on cardiac intensive care unit (CICU) or cardiac surgical ward]

• written informed consent.

The exclusion criteria were:

• patients undergoing emergency cardiac surgery

• patients prevented from having blood and blood products according to a system of beliefs (e.g. Jehovah’s Witnesses)

• patients with congenital or acquired platelet, red blood cell or clotting disorders

• patients with ongoing or recurrent sepsis

• patients with critical limb ischaemia

• patients unable to give full informed consent for the study (e.g. learning or language difficulties)

• patients already participating in another interventional research study.

Changes to study eligibility criteria after commencement of the study

In April 2009 (before starting recruitment to the study), two exclusion criteria were removed:

• patients with a critical carotid artery stenosis (> 75%)

• patients with flow limiting (> 70% luminal stenosis) coronary artery disease not undergoing complete revascularisation.

These exclusion criteria were included originally on the basis of the exclusion criteria used in the pilot study for this trial. 26 The pilot study used different thresholds, notably a lower haemoglobin threshold of 7 g/dl for the restrictive group. At the time of designing the pilot study it was felt that, because patients entering the pilot study could potentially experience haemoglobin levels as low as 7 g/dl, these exclusion criteria were needed to avoid non-adherence by intensivists, who might consider such patients to be more at risk of experiencing ischaemic adverse effects. TITRe2 used the higher haemoglobin level of 7.5 g/dl for the restrictive threshold and this threshold was already used routinely at some centres for all patients. Therefore, after discussing these exclusion criteria again, the study team believed they were not necessary for TITRe2.

In August 2010, the previously stated upper age limit of 80 years for the inclusion of participants was removed. This decision was a result of feedback from sites that they did not consider older age to be a contraindication for randomisation to the study and that the exclusion would substantially limit the pool of eligible patients for the study. After removal of this criterion surgeons were still able to refuse to include patients aged > 80 years on a case-by-case basis.

Primary outcome

The primary outcome was a binary composite outcome of any serious infectious (sepsis or wound infection) or ischaemic event [permanent stroke, myocardial infarction (MI), gut infarction or AKI] in the first 3 months after randomisation. The qualifying events listed in Table 1 were included; the table also describes the manner in which each qualifying event was verified.

Events occurring after discharge only contributed to the primary outcome if the potentially qualifying event resulted in admission to hospital or death. Wound infection identified as a result of adding post-discharge information was the only exception to this rule. For example, information ascertained using the additional treatment, serous discharge, erythema, purulent exudate, separation of deep tissues, isolation of bacteria, and stay duration as inpatient (ASEPSIS) post-discharge surveillance assessment questionnaire (see Table 1), when added to the ASEPSIS score derived for the index admission, sometimes resulted in a total ASEPSIS score for the index admission that was > 20. Other suspected infectious events treated in the community that did not cause readmission to hospital were not recorded as they could not be validated and are less serious than perioperative infections.

Events suspected to qualify for the primary outcome but that were not supported by objective evidence were referred to an independent adjudication committee. In practice, the adjudication committee only considered suspected MIs because documentary objective evidence for sepsis, stroke, AKI and gut infarction was verified by research nurses at the co-ordinating centre who were blinded to the random allocation. The adjudication committee consisted of a cardiac surgeon, cardiologist and anaesthetist who were blinded to allocation and each other’s assessments. They were required to classify a suspected MI as definite or not based on participant’s medical history, echocardiograms (ECGs) (both pre-operatively and at the time of the suspected MI) and troponin levels at the time of the suspected MI. Agreement between at least two of the three specialists was required to reach a final adjudicated classification.

Death was not included as a component of the primary composite outcome because, if death occurred following a qualifying event, the event would precede death itself. Deaths that occurred for other reasons were not hypothesised to increase because of red blood cell transfusion.

Secondary outcomes

All secondary outcomes were collected in the time between randomisation and 3-month follow-up, unless otherwise stated.

• Units of red blood cells and other blood components [fresh frozen plasma (FFP), platelets, cryoprecipitate, activated factor VI (NovoSeven, Novo Nordisk) and Beriplex^® (CSL Behring UK Ltd)] transfused during a participant’s hospital stay. Red blood cells transfused pre-randomisation (either intraoperatively or postoperatively but prior to randomisation) were collected and described separately. However, it was only possible to collect information about other blood components transfused over the pre-randomisation and post-randomisation periods combined.

• Occurrence of an infectious qualifying event, defined as sepsis or wound infection.

• Occurrence of an ischaemic qualifying event, defined as permanent stroke, MI, AKI or gut infarction.

• Quality of life measured using European Quality of Life-5 Dimensions-3 Level (EQ-5D-3L),31 assessed pre-operatively and at 6 weeks and 3 months post randomisation.

• Duration of intensive care unit (ICU) or high-dependency unit (HDU) stay; calculated as the total time between randomisation and discharge from the cardiac unit that the participant was in either the CICU, HDU or general ICU wards, including any periods of readmission to that area.

• Duration of hospital stay, calculated as the time between randomisation and discharge from the cardiac unit.

• Significant pulmonary morbidity, comprising initiation of non-invasive ventilation [e.g. continuous positive airway pressure (CPAP) ventilation], reintubation/ventilation or tracheostomy.

• All-cause mortality.

• Health and Personal Social Services resource use and their costs.

(Durations of ICU, HDU and hospital stay were originally specified as ‘postoperative’. We specified randomisation as the time origin for these durations in the analysis plan, for consistency with the primary and other secondary outcomes.)

Changes to study outcomes after commencement of the study

The following changes were made to study outcomes after the trial had commenced.

• In April 2009, before starting recruitment, there were some amendments made to the definitions of infectious and ischaemic events that qualified for the primary outcome.

• In March 2010, an amendment was made to include troponin T in addition to troponin I in defining MI. This amendment was required after discovering that some participating centres habitually used troponin T rather than I. In addition, as part of this change, the troponin threshold for MI that was previously stated was removed; the decision was made, instead, to collect the highest troponin reading for all participants with suspected MI and to adjudicate suspected MIs (see Primary outcome).

• In March 2011, the secondary outcome ‘significant pulmonary morbidity’ was added. This was initially named transfusion-associated circulatory overload and then subsequently renamed. The outcome was added because information from the haematology community had highlighted pulmonary morbidity as a potentially important outcome for patients receiving blood transfusions. The outcome was defined with respect to data already being collected on the study case report forms (CRFs) before the amendment so that the outcome could be identified consistently across the entire duration of the trial.

• Furthermore, in March 2011, ‘A&E [accident and emergency] admission’ was removed from the primary outcome as qualifying event. This change arose from discussion with clinicians on the Trial Steering Committee (TSC) who agreed that a participant experiencing any element of the primary outcome would be admitted to hospital if they attended the emergency department (ED) within the follow-up period.

Adverse events

Expected adverse events (AEs) were specified in the study protocol and captured via the study CRFs, both for the post-operative in-hospital period (serious and non-serious), and at the 3-month follow-up (serious only).

A serious adverse events (SAE) is any untoward medical occurrence that either results in death, is life-threatening, requires in-patient hospitalisation or prolongation of existing hospitalisation, results in persistent or significant disability or incapacity, or results in a congenital anomaly/birth defect.

All other AEs were considered unexpected, any such events satisfying one or more criteria for classification as serious were recorded in detail on purpose-designed SAE forms. Unexpected SAEs were coded using the Medical Dictionary for Regulatory Activities version 14.1 (MedDRA; McLean, VA, USA) independently by two research nurses blinded to randomised allocation. Any discrepancies were resolved by a consultant cardiac surgeon also blinded to allocation.

Non-adherence with randomisation protocol

Non-adherence with the randomisation protocol was defined as any of the following:

• Participant did not meet one or more of the pre-consent study eligibility criteria but was consented into the study. Any randomised participant to whom this applies was classified as ‘randomised in error’ and excluded from the analysis population.

• Participant consented and met the post-consent inclusion criteria (i.e. haemoglobin dropped below 9 g/dl) but was not randomised. Any randomised participant to whom this applies was not randomised and, therefore, was excluded from the analysis population.

• Participant did not meet the post-consent eligibility criteria (i.e. haemoglobin did not drop below 9 g/dl) but was randomised. Any randomised participant to whom this applies was classified as ‘randomised in error’ and excluded from the analysis population.

• Participant was randomised more than 24 hours after meeting the post-consent inclusion criteria (i.e. randomised more than 24 hours after haemoglobin dropping below 9 g/dl). Any participant to whom this applies was classified as non-adherent with the randomisation protocol, but was included in the analysis population.

Non-adherence with transfusion protocol

Measuring and assessing adherence with the transfusion protocol was identified as a critical element of the study owing to the assumptions about adherence made in the sample size calculation. 33 The Data Monitoring and Ethics Committee (DMEC) also highlighted the importance of non-adherence, as they had concern that doctors might otherwise make transfusion decisions in different ways in the two groups. For example, a decision to transfuse in the liberal group could be delayed up to 24 hours without contravening the protocol and such behaviour would have been missed if data about haemoglobin levels measured during this period, their times and consequent actions had not been recorded.

Two types of non-adherence were defined: (1) a participant received a red blood cell transfusion outside of protocol (‘extra’ transfusion) and (2) a participant was not given a red blood cell transfusion that, according to the protocol, should have been given (‘withheld’ transfusion). Adherence was assessed for the period from randomisation to hospital discharge so multiple instances of non-adherence could be documented for a participant. If a participant withdrew or had their treatment according to their allocation discontinued, adherence after the time of withdrawal/discontinuation was not assessed. For both of the above types of non-adherence, instances were classified into mild, moderate or severe (Table 3). Non-adherence was classified as severe only if the non-adherent instant changed the participant’s overall classification as transfused or not.

In addition to describing the amount of non-adherence, work has been done to further describe and characterise non-adherence, including:

• characteristics of each instance of non-adherence (including reasons, haemoglobin levels and timing) have been described

• logistic regression models were fitted to identify predictors of non-adherence

• non-adherence trends both by centre and over the course of the trial have been described.

Comparisons of outcomes

All outcomes were analysed using mixed-effects regression models, adjusting for all factors included in the cohort minimisation: operation type as a fixed effect and centre as a random effect (or a shared frailty term in time-to-event models). The primary outcome and other binary outcomes (numbers of participants experiencing infectious or ischaemic events, any transfusions of red blood cells and non-red blood cell products and significant pulmonary morbidity) were analysed using logistic regression, with treatment estimates presented as odds ratios (OR) and 95% confidence intervals (CI). For the analysis of the transfusion of any red blood cells, treatment estimates were analysed using unadjusted logistic regression, with results presented as a risk ratio (RR) and 95% CI, as the OR proved difficult to interpret and an adjusted model did not converge. Time-to-event outcomes were analysed using Cox proportional hazards models and treatment estimates presented as hazard ratios (HR) and 95% CI. Durations of ICU/HDU stay and hospital stay were censored at the time of death if the participant died before discharge from hospital. All-cause mortality was censored at the time of last follow-up for survivors. A secondary analysis of the primary outcome, analysing the time to first occurrence of the primary outcome, was also undertaken using a Cox proportional hazards model, censoring at the time of last follow-up or death.

Longitudinal data (EQ-5D-3L scores) were analysed using mixed-effects mixed-distribution models;35 this method was used because the distribution of the data was non-monotonic, with many participants scoring perfect health. Both types of score (utility and visual analogue scores) were dichotomised into less than perfect health compared with perfect health. There were two-parts to each fitted model: (1) an occurrence model, a logistic regression model for the occurrence of less than perfect health compared with perfect health, and (2) an intensity model, a log-linear model for the score, conditional on a non-perfect health score. Correlated participant-term random effects (for occurrence and intensity) were included in each model to allow for the repeated measures. Separate parameter estimates were incorporated into models for the mean baseline response across both treatment groups and for each treatment post intervention, avoiding the necessity to either exclude cases with missing baseline measures or to impute missing baseline values. A time by allocation interaction (post intervention) was added to each of the models; an overall treatment effect is reported unless the interaction was statistically significant at the 10% level, in which case separate treatment effects at each post-intervention time are given.

Safety data

The AEs and SAEs were described by allocated treatment but no formal comparisons made.

Subgroup analyses

Pre-planned subgroup analyses were specified because of clinical opinion that transfusion decisions should be influenced by patients’ characteristics, notably that ‘at-risk’ patients should be transfused at a different threshold. The subgroups defined in the protocol were: operation type (isolated CABG vs. other operation types), age (< 75 years vs. ≥ 75 years), pre-operative diagnosis of diabetes (none vs. diet, oral medication or insulin controlled), pre-operative diagnosis of lung disease (none vs. chronic pulmonary disease or asthma), pre-operative renal impairment [estimated glomerular filtration rate (eGFR) > 60 ml/minute vs. eGFR ≤ 60 ml/minute], sex (males vs. females) and ventricular function (good vs. moderate or poor). Such analyses were implemented by adding a relevant treatment allocation by subgroup interaction term into the primary outcome model; the hypothesis for all subgroup analyses was that there would be no interaction. The pre-operative renal impairment subgroup analysis was defined in the study protocol as pre-operative creatinine ≤ 177 µmol/l versus creatinine > 177 µmol/l. However, during the course of the trial, use of pre-operative creatinine for risk stratification was superseded by estimated eGFR and, therefore, the subgroup analysis for renal impairment was based on eGFR as described above (this change was not covered by a protocol amendment).

Sensitivity analyses

A number of sensitivity analyses were pre-specified for the primary outcome in the analysis plan, although such analyses were not specified in the study protocol.

1. Examining treatment effect estimates for the primary outcome by site, ordering sites by rates of severe non-adherence with the transfusion protocol. This was implemented by a forest plot displaying site-specific treatment estimates. It provided a way of assessing the effect of non-adherence on the overall treatment estimate for the primary outcome, without excluding non-adherent participants. (We considered that an analysis excluding non-adherent participants would be inappropriate because it would be very likely to be biased as non-adherent participants were hypothesised to be the sicker participants in the restrictive group and the healthier participants in the liberal group.) As non-adherence represents a dilution of the allocated intervention, we hypothesised that the treatment effect would tend towards the null with increasing non-adherence.

2. Excluding all events that occurred in the first 24 hours after randomisation. The rationale was that such events could have an onset that actually preceded randomisation and, hence, be unrelated to the intervention. Therefore, we hypothesised that the treatment effect would tend away from the null with exclusion of these events.

3. Excluding participants who were transfused before randomisation. The rationale was that transfusions before randomisation, expected to occur with similar frequency in both groups, would dilute any effect of a difference between groups in the number of transfusions after randomisation. Therefore, we hypothesised that the treatment effect would tend away from the null with exclusion of pre-randomisation transfusions.

4. In collecting AKI data it became apparent that prospective data collection by research nurses in centres failed to identify AKI events that were apparent from routinely collected serial creatinine data. We attribute this discrepancy to differences between centres in the ‘baseline’ creatinine value used to define AKI, which can be confusing to implement as the specified creatinine rise should occur in a 48-hour period. 30,36 However, highest daily creatinine levels were recorded separately, so the following sensitivity analyses were planned:

   • excluding AKI events when the clinical diagnosis was not verified by routinely recorded creatinine levels. This analysis would exclude potentially ‘false’ AKI events, although AKI events in these participants may have been ‘true’ events classified on the basis of urine output (which was not documented routinely).

   • including additional AKI events when the participant was reported not to have had AKI according to clinical judgement but when highest daily creatinine levels supported a diagnosis of AKI. This analysis would include AKI events that were missed, assuming that the creatinine levels recorded for usual hospital care were accurately transcribed on to the CRF.

   Assuming that false AKI events would arise in proportion to the incidence of true AKI events, they would not bias the treatment effect. Therefore, we hypothesised that the effect in the first analysis would reduce precision but not shift the estimate predictably either towards or away from the null. Similarly, assuming that missed AKI events would arise in proportion to the incidence of true AKI events, they would also not bias the treatment effect. Therefore, we hypothesised that the effect in the second analysis would increase precision but not shift the estimate predictably.

5. Including only ‘serious’ primary outcome events, defined as either stroke, MI, gut infarction, AKI stage three events, pre-discharge sepsis plus organ failure [MI, stroke, laparotomy for gut infarction and one or more of reintubation, acute respiratory distress syndrome (ARDS), low cardiac output and/or tracheostomy] and/or post-discharge sepsis that required hospital readmission. This analysis arose from the pre-planned interim analysis that showed a higher primary outcome event frequency than was anticipated when the study was designed, with a large majority of qualifying events arising from sepsis and AKI, which were considered to be clinically less serious. We considered that this sensitivity analysis would better reflect our original intention in formulating the composite outcome and the outcome events that were included in the observational analysis, which led to the superiority hypothesis for the trial. This analysis would necessarily have less precision but we did not have a strong hypothesis about the way in which the treatment effect might be affected. If transfusion were to have the same effect on more and less serious events, the treatment effect should be unaltered; if transfusion were to have a differential effect on more and less serious events, the treatment effect should be moved towards or away from the null in a manner consistent with the differential effect.

Post-hoc analyses

In addition, a secondary post-hoc analysis of severe in hospital events was performed, which involved refitting the primary outcome model with an outcome of death, severe sepsis [as defined in sensitivity analysis (e) above], ARDS, tracheostomy, low cardiac output, MI, AKI stage three, gut infarction and/or stroke. This analysis was performed because it was judged to be of key interest to hospital-based clinicians caring for patients. As for analysis (e) above, it would necessarily have less precision but we did not hypothesise that the treatment effect would be moved towards or away from the null.

Two further post-hoc sensitivity analyses were carried out for the secondary outcome of mortality; these comprised analyses (b) and (c) above (i.e. excluding deaths within 24 hours of randomisation and participants transfused before randomisation). These were suggested during the peer review process, on account of the seriousness of the outcome, and we agreed that they were worthwhile. Our (post hoc) hypotheses were the same as that for the corresponding sensitivity analyses of the primary outcome. These were the only additional analyses requested in this way, the decision to perform them was made without knowing the results and the results are fully reported.

Observational analyses

Three observational analyses were pre-specified in the study protocol.

1. Estimating the relationship between the number of red blood cell units transfused and the risk of mortality and morbidity, stratified by trial arm.

2. Investigating the relationship between percentage decline in haemoglobin from the pre-operative level and the risk of primary and secondary outcomes, taking into account the number of red blood cell units transfused.

3. Investigating whether or not red blood cell age (i.e. time since donation and processing) is associated with the risk of primary and secondary outcomes, achieved by linking batch numbers of all red blood cells transfused to a blood bank database and determining the age of each unit donation and transfusion dates.

For the purposes of these observational analyses, a composite outcome of the trial primary outcome or death was defined in the statistical analysis plan (SAP). In addition, all red blood cell units transfused or haemoglobin levels recorded after the time of the first occurrence of the primary outcome (or censoring) were excluded to ensure the relevant exposure occurred before the outcome. (More complex methods to deal with this issue were outlined in the SAP but these have not been attempted owing to the complexity of the analyses.) For all three analyses, pre-operative and intraoperative characteristics and trial outcomes were described by exposure [i.e. any red blood cells vs. no red blood cells, minimum haemoglobin < 7.5 g/dl vs. ≥ 7.5 g/dl, and transfusion of any red blood cells aged over 21 days old (median age) vs. only younger blood (< 21 days) vs. no red blood cell transfusions].

For analyses (b) and (c), the exposure definitions differ slightly from those used in the protocol/SAP. With respect to analysis (b), the protocol stated that haemoglobin would be defined in terms of percentage decline; however, exploratory analyses suggested this was not sensible (e.g. a participant with pre-operative haemoglobin 12 g/dl and post-randomisation haemoglobin 6 g/dl would be treated in the same way as a participant with pre-operative haemoglobin 18 g/dl and post-randomisation haemoglobin 9 g/dl, as the percentage decline is 50% in both cases) and that it would be more informative to include both pre-operative and post-randomisation haemoglobin levels in any analysis model. For analysis (c), various methods of defining age of blood were described in the SAP (using the age of the ‘oldest’ red blood cell unit given, the mean age of all red blood cells, the use of any red blood cells more than 14 days old, the number or percentage of red blood cells given over 14 days old, the use of red blood cells older than the median age of all red blood cells transfused) and it was stated that the age of the oldest unit would be used as the primary analysis. Owing to the large volume of missing data for age of blood, it was decided instead only to provide descriptive analyses by the receipt of any red blood cells older than the median age.

For parts (a) and (b), further analyses have been undertaken. Univariate analyses exploring the relationship between exposures and the outcome were performed. Two separate adjusted models [one for analysis (a) and one for analysis (b)] were then fitted adjusting for the following, if found to be potential confounders: operation type, centre (as a random effect), European System for Cardiac Operative Risk Evaluation (EuroSCORE), age, sex and pre-randomisation red blood cell transfusions [for analysis (a) only]. A model building strategy was used whereby variables were sequentially added to the model, at each step including the variable that improved the model fit the most (as determined by a likelihood ratio test). Variables were included in the model if they were (1) associated with both the exposure and the outcome but did not lie on the causal pathway between the exposure and outcome, and (2) significantly contributed to the relevant multivariate model (defined by a likelihood ratio p < 0.05 or modifying the effect estimate by greater than 10%). If pairs of variables were considered to be collinear or strongly related (e.g. EuroSCORE and age), only one of the pair was included. In addition, interaction terms were included in models if significant at the 5% level. The parameterisation of the exposure variable (e.g. continuous linear, continuous including additional power terms, ordinal categorical or binary) was explored using fractional polynomial models and likelihood ratio tests to compare nested models. Marginal plots were used to describe interactions between continuous and categorical covariates graphically. Models were refitted separately within each randomised group. Finally, instrumental variable (IV) methods were used to estimate the associations of interest free from confounding, separately for analyses (a) and (b); models used the multiplicative generalised method of moments estimation (the ivpoisson command in Stata).

Meta-analysis

A meta-analysis was performed analysing mortality from TITRe2 and all other RCTs that have compared liberal and restrictive red blood cell transfusion strategies in patients undergoing cardiac surgery. This analysis was undertaken to place the findings of TITRe2 in the context of the evidence base. Eligible RCTs24–26,37,38 were identified from a previous review of RCTs comparing restrictive versus liberal transfusion thresholds12 and an on-going review comparing RCT and observational evidence about the effects of red blood cell transfusion in cardiac surgery patients. 39 The previous Cochrane review searched multiple databases including the Cochrane Injuries Group’s Specialised Register, the Cochrane Central Register of Controlled Trials, MEDLINE, EMBASE and ISI Web of Science (both the Science and Conference Proceedings Citation Indices). The review included RCTs with a concurrent control group in which participants were assigned to groups with different transfusion triggers or thresholds, for which the thresholds were defined by a haemoglobin or haematocrit level that a participant had to reach before a red blood cell transfusion could be administered. For the purposes of the current meta-analysis, we included RCTs identified from either review that were deemed to have taken place in the context of cardiac surgery. Therefore, we included RCTs with different group-specific transfusion thresholds to those used in TITRe2 and which included all participants, irrespective of whether or not the liberal threshold was breached (i.e. without the post-operative eligibility criterion adopted in the TITRe2 trial). When writing the SAP, we also intended to perform a meta-analysis for the primary outcome; however, outcomes were too dissimilar between included RCTs. The meta-analysis was performed using standard meta-analysis methods for binary outcomes with a random effects model.

Missing data

Missing data are indicated in all of the tables. Rules for imputing missing data were outlined in the analysis plan, dependent on the level of missing data. However, the majority of outcomes had levels of missing data below the defined thresholds in the plan (5% for outcomes measured at one time point and 20% for longitudinal data) and imputation methods were not generally used. The first exception was for the infectious events secondary outcome (5.6% missing), whereby separate estimates were made prior to hospital discharge and overall. A second exception was the in-hospital component of the ASEPSIS score from which wound infection events were identified; this was one of the rarer components of the primary outcome but the in-hospital component of the score was the outcome data item that was most likely to be missing. If the in-hospital ASEPSIS score was missing, the participant was assumed not to have had a serious wound infection if the following criteria were met: participant did not have antibiotics for suspected wound infection prescribed in hospital and follow-up was completed, and the participant reported no problems with the healing of chest, leg and/or arm wound up to 3 months after the operation.

Significance levels

For hypothesis tests, two-tailed p-values of < 0.05 were considered statistically significant, with the exception of tests for interactions between group and time in longitudinal models when a 10% significance level was used. Likelihood ratio tests were used in preference to Wald tests. No formal adjustment was made for multiple testing. When interpreting the results, consideration has been given to the number of tests performed and the consistency, magnitude and direction of estimates for different outcomes. 40 All data management and analyses were performed in SAS version 9.3 (SAS Institute Inc., Cary, NC, USA) or Stata version 12.1 or 13.1 (StataCorp LP, College Station, TX, USA).

Aims and objectives

The economic evaluation aimed to estimate the cost-effectiveness of the restrictive compared with the liberal haemoglobin transfusion threshold as compared in TITRe2. Our main objective was to estimate the incremental cost and the incremental cost-effectiveness of the restrictive compared with the liberal haemoglobin transfusion threshold after cardiac surgery.

Economic evaluation methods overview

A cost–utility analysis was conducted, with outcomes measured using the EQ-5D-3L. The restrictive haemoglobin threshold was considered as cost-effective if the incremental cost-effectiveness ratio (ICER) fell below £20,000, which is generally considered as the threshold at which the National Institute for Health and Care Excellence (NICE) considers an intervention to be cost-effective. 41 Good practice guidelines on the conduct of economic evaluations were followed for the economic evaluation. 42–44 Table 4 summarises the methods for the economic evaluation, with further details provided in the text following the table.

Form of analysis and primary outcome measure

The primary outcome measure for the economic evaluation was QALYs, as advocated by NICE. 44 This outcome combines quantity and quality of life into a single measure. Our evaluation took the form of a cost–utility analysis in which the difference in mean costs between the two transfusion threshold groups is divided by the difference in mean QALYs between the two groups to calculate an ICER and, specifically, the incremental cost per QALY gained by switching from using a liberal threshold to using a restrictive threshold.

Perspective

The primary perspective of the evaluation was that of the UK NHS and Personal Social Services, as recommended by NICE. 44 However, data were collected on some types of non-NHS costs including expenditure incurred by a participant when travelling to hospital. We planned to include these costs in a wider perspective in a sensitivity analysis, if resource use for these non-NHS costs differed between the trial groups. The perspective for outcomes was that of the participants undergoing treatment.

Time horizon

A within-trial analysis, taking a 3-month time horizon, was conducted. It was anticipated that all major resource use would occur within this timeframe and, therefore, be captured. The start of our analysis was from the point of surgery. Surgery was chosen as the time origin, rather than the point of randomisation as was the case with the analysis of effectiveness, in order to capture the resources that would be required for the intervention from a decision-maker’s perspective, that is, to include all relevant costs (and effects) involved in delivering an intervention. Our time horizon continued until 3 months postoperatively. Ideally, the time point for baseline costs and outcomes should be the same; however, the EQ-5D-3L was collected pre-operatively whereas detailed resource use collection began on the day of surgery.

Population

Our base-case analysis included all participants randomised into the trial except those randomised in error, which is consistent with the main effectiveness analyses. Analyses were performed on an ITT basis.

Collection of resource use and cost data

Resource use data were collected on all significant health service resource inputs for the trial participants up to the point of the 3-month follow-up. The main resource use categories that were costed are listed in the first column of Table 5, along with the sources of information for both the resource use and unit costs. Costing decisions (such as resource use assumed for complications) were made without knowledge of the allocation of participants to trial groups.

Attaching unit costs to resource use

Unit costs for hospital and community health-care resource use were largely obtained from national sources, for example, NHSBT price lists for blood products, the National Schedule of Reference Costs for ICU, HDU and cardiac ward costs, MRI and CT scans and many complications, and Unit Costs of Health and Social Care for community costs. 45,46,49 Resources were valued in 2012/13 pounds sterling (£); if any unit costs were in pre-2012/13 prices, they have been inflated to 2012/13 using the Hospital and Community Health Services (HCHS) inflation index. 49 Costs of drugs given in hospital were taken from the electronic marketing information tool (eMIT)47 when possible, which provides the reduced prices paid for generic drugs in hospital; other drug costs were taken from the British National Formulary (BNF). 48 Tables 63 and 64 in Appendix 3 lists all the medications and their costs used for the trial; further details on all unit costs and their source can be found in Appendix 3, Unit costs and resource use assumed for complications.

Measurement of health-related quality of life and quality-adjusted life-years

Missing data

We first summarised descriptively the volume of missing data for both resource use and EQ-5D-3L scores, which showed that 2.5% of resource use data were completely missing: 2.4% in the restrictive group and 2.5% in the liberal group. Overall, 10.7% of EQ-5D-3L scores were missing across the three time points (pre-surgery, 6 weeks and 3 months); 10.9% in the restrictive group and 10.5% in the liberal group. Although the level of missing data on resource use sounds small, because there are a large number of resource use variables for the trial, any simple methods to deal with the missing data would work poorly. For instance, using complete case analysis would leave only 61% of participants remaining for analysis. Multiple imputation was used to handle this missing data.

When data were partially missing, for example for linked questions for which only the first part was answered, mean imputation was used to handle such missing data. This occurred when a resource use question was in two parts: if participants were asked to respond yes/no to whether or not they used a particular resource (e.g. if they were readmitted to hospital) and then if yes, to record further details on the volume of resource use (e.g. the number of days they were readmitted, or the number of visits). Further details of mean and multiple imputation are described next.

Adjustment for baseline utility

Given that baseline utility directly contributes to QALY calculations, it is important to control for any potential imbalances in baseline utility in the estimation of the mean difference in QALYs between treatment groups, to avoid introducing bias. 53 Regression adjustment also allows for regression to the mean and increases precision. Therefore, we adjusted our QALYs for baseline EQ-5D-3L. For each of the five imputed datasets, we regressed total QALYs on treatment group and baseline EQ-5D-3L and used the Stata command, mi estimate, to combine the five imputed datasets, and used Rubin’s Rule to combine the standard errors (SEs) across the five imputations. This provided an estimate of the QALY difference and its SE between the trial groups, adjusted for baseline EQ-5D-3L. The Stata command nlcom was then used to combine regression coefficients to obtain the mean QALYs in each trial group, adjusted for baseline EQ-5D-3L.

Within-trial statistical analysis of cost-effectiveness results

Most of the cost-effectiveness analyses were conducted in Stata version 12; some of the graphs and unit cost calculations were conducted in Microsoft Excel^® 2010 (Microsoft Corporation, Redmond, WA, USA).

Initially, resource use, costs and health-related quality of life were summarised using means, SDs and SEs of means, using both the central limit theorem and bootstrapping. ICERs were derived from the average costs and QALYs gained in each trial group, producing an incremental cost per QALY gained by implementing a restrictive threshold in place of a liberal threshold. Non-parametric bootstrapping of costs and QALYs was then used to quantify the degree of uncertainty around the ICER. Bootstrapping was used to avoid making parametric assumptions.

A thousand bootstrap samples were drawn for each of the five imputed datasets. For each bootstrap sample for each imputation, total costs were regressed on treatment group and total QALYs were regressed on treatment group and baseline EQ-5D-3L. The mean cost difference between the groups (restrictive minus liberal) was calculated, as well as the mean QALY difference between the groups adjusted for baseline EQ-5D-3L.

These 5000 bootstrap replicates of the mean difference in costs and QALYs between the groups were used to represent graphically the uncertainty around the ICER on the cost-effectiveness plane. In order that the points could be seen, only 1000 replicates were plotted (200 replicates for each of the five imputations).

For each of the five imputations, the mean and SD of the 1000 cost and QALY differences were estimated. These SDs are SDs of a column of means, so are actually SEs. Rubin’s Rule was then used to combine the SEs across the five imputed datasets. CIs around the cost and QALY differences were then generated based on these SEs rather than SEs based on parametric methods.

Results are expressed in terms of a cost-effectiveness acceptability curve (CEAC), which indicates the likelihood that the restrictive threshold is cost-effective for different levels that health-care decision-makers are willing to pay for health gain. All 5000 bootstrap replicates were used to generate the CEAC. The restrictive threshold would be considered as cost-effective if the ICER falls below £20,000; however, the ICERs and CEACs presented would allow decision-makers to assess cost-effectiveness at a willingness-to-pay threshold of their choice.

Sensitivity analysis

One-way sensitivity analysis was used to investigate the impact on the results of the cost and cost-effectiveness analyses when varying key parameters or major cost drivers and also to investigate the impact of uncertainty on the cost-effectiveness results. Factors that were examined in the sensitivity analysis for costing were:

• varying the unit costs of treatment of complications, ward stays, reoperations, expensive drugs, oral versus intravenous drug administration and the source of medication unit costs (BNF vs. eMIT) (see Appendix 3, Sensitivity analyses around unit costs)

• conducting the costing from the point of randomisation rather than the point of surgery as undertaken in the baseline analysis (further details in Appendix 3, Costs from randomisation)

• exploring the impact of any high-cost participants (outliers) if the cost data are skewed.

Non-NHS costs were also considered and analyses were conducted to determine whether or not these differed between the trial groups and hence whether or not there was a need to conduct a sensitivity analysis from a wider societal perspective (instead of a NHS and Personal Social Services perceptive).

For the sensitivity analysis on outcomes, we varied the assumptions for calculating QALYs; the alternative strategies examined were:

• not adjusting for baseline utility

• exploring the use of the last observation carried forward until death rather than assuming utility changes linearly until death

• using the date of completion of the 6-week EQ-5D-3L rather than assuming it is completed at exactly 6 weeks

• calculating QALYs from the point of randomisation rather than surgery.

Finally, we carried out a sensitivity analysis examining life-years gained as a secondary outcome measure in the economic evaluation. Previous economic evaluations conducted by the authors have often found that EQ-5D-3L scores are similar across trial groups. 54 Such a finding could reflect reality but could also be a function of the 3-level version of the EQ-5D-3L not being sufficiently sensitivity to changes in quality of life, or that quality-of-life improvements arise before EQ-5D-3L measurements (failing to capture periods of lower quality of life), especially following SAEs.

Subgroup analysis

Subgroup analyses were conducted to investigate whether or not cost-effectiveness results varied between participant subgroups. The pre-specified subgroups used for the effectiveness analyses were used for the cost-effectiveness subgroup analyses:

• operation type (isolated CABG vs. other operation types)

• age at operation (< 75 years vs. ≥ 75 years)

• pre-operative diagnosis of diabetes (none vs. diet, oral medication or insulin controlled)

• pre-operative diagnosis of lung disease (none vs. chronic pulmonary disease or asthma)

• pre-operative renal impairment (eGFR ≤ 60 ml/minute vs. eGFR > 60 ml/minute)

• sex (males vs. females)

• pre-operative ventricular function (good vs. moderate or poor).

The impact of subgroups was evaluated using ordinary least squares regression predicting total costs and QALYs, conditional on treatment group, subgroup and an interaction between treatment group and subgroup (and baseline EQ-5D-3L for QALYs only). A Bonferroni adjustment was made to allow for the multiple tests conducted across the seven subgroups and two variables; statistical significance was therefore evaluated at the 0.0036 level.

Non-adherence with the randomisation protocol

There were nine participants consented to the trial who were later found to be ineligible, although none of these participants were randomised (Table 16). Of the 1464 non-randomised participants, 176 (12.0%) met the post-consent eligibility criterion (haemoglobin < 9 g/dl) but were not randomised. All of the 2003 randomised participants breached the 9 g/dl threshold. However, randomisation was delayed (i.e. occurred later than 24 hours after the threshold breach occurred) for 65 participants (3.2%) and these instances were classified as non-adherent with respect to the randomisation protocol.

Non-adherence with the allocated transfusion threshold

Non-adherence with the allocated transfusion threshold is described in Table 17. There were 1813 deviations from the protocol occurring in 37.6% of participants; 635 deviations in 30.0% of participants in the restrictive group and 1178 deviations in 45.2% of participants in the liberal group. As anticipated, extra transfusions (i.e. given outside of protocol) were more common in the restrictive group (573 in 27.3% of participants vs. 161 in 10.7% of participants) and withheld transfusions were more common in the liberal group (1017 in 38.9% of participants vs. 62 in 5.5% of participants). Approximately one-sixth of all deviations were classified as severe; 24.8% of extra transfusions and 11.1% of withheld transfusions. Therefore, severe protocol deviations were more common in the restrictive group than the liberal group (186 in 9.7% of participants vs. 116 in 6.2% of participants).

Characteristics of each instance of non-adherence are reported in Table 18. Extra transfusions tended to be given either for the clinical reasons listed on the CRF (36.5%) or for ‘other’ reasons (42.7%), which were generally clinical reasons not listed as specific options on the CRF. The most common reason for a withheld transfusion was ‘oversight/error’ (67.2%). Extra transfusions tended to occur earlier than withheld transfusions and were more likely to occur overnight. There were no clear trends in time of year for either type of non-adherence.

Separate logistic regression models were fitted for (1) extra transfusions and (2) withheld transfusions to identify any characteristics (both at an adherence level and participant level) that predicted non-adherence (Table 19). The odds of an extra transfusion reduced by 3% with each post-operative day and reduced by 22% at weekends. However, the odds of a withheld transfusion increased by 3% with each post-operative day and increased by 79% at weekends. Both types of non-adherence were much more likely in the ICU than on the ward. Centre recruitment rate was important for predicting both types of non-adherence, which was most common in relatively slow recruiting centres (3–4 participants per month). Participants having more complex operation types (CABG and valve and valve-alone surgery) were more likely to have an extra transfusion, as were participants transfused pre-randomisation. Participants with a longer period of time between operation end and randomisation were more likely to have a withheld transfusion, and the odds of a withheld transfusion decreased by 1% with each year of age.

Adherence by centre is given in Figure 9, which demonstrates wide variation between centres with no obvious relationship to total recruitment or average recruitment rates.

FIGURE 9.

Adherence by centre. (a) Percentage of participants with any severe/non-severe non-adherence by centre. (b) Percentage of participants having extra and withheld transfusions. Sites are ordered by the average number of participants recruited per month.

Non-adherence was monitored carefully over the course of the trial and various measures were put in place to try and improve non-adherence rates. 33 Some of these measures are described in Table 20.

There was no improvement in non-adherence rates over the course of the study despite these measures (Figure 10). In the early stages, rates of non-adherence fluctuated somewhat but by the time that half of the participants had been recruited, rates remained fairly constant.

FIGURE 10.

Change over time in proportions of non-adherent participants. Lines represent cumulative non-adherence rates; triangles represent the number of centres in the study.

Primary analysis

The primary outcome occurred in 35.1% of participants in the restrictive group and 33.0% in the liberal group (Table 21). This difference was not statistically significant, OR 1.11 (95% CI 0.91 to 1.34; p = 0.30). The most common element of the primary outcome was sepsis (21.6%), followed by AKI (13.2%) and wound infection (5.4%); all other components were relatively rare (< 2%). There were 27.1% participants who experienced the primary outcome before hospital discharge and 10.8% after hospital discharge (some participants experienced qualifying events both in hospital and after discharge). The small excess of primary outcome events in the restrictive group appeared to be mainly driven by AKI events. Most of the AKI events occurred before hospital discharge (96.6%) and AKI events had similar frequencies in each of the three AKI stages (34.4% stage one, 28.6% stage two and 37.1% stage three).

Most participants experiencing the primary outcome encountered their first component event in the first 10 days after randomisation (Figure 11). A planned secondary analysis using a Cox proportional hazards model gave HR 1.09 (95% CI 0.93 to 1.27; p = 0.29) which is very similar to the OR obtained from logistic regression in Table 21.

FIGURE 11.

Kaplan–Meier estimates of time from randomisation to the primary outcome. The HR is not adjusted for cardiac procedure (as was planned) because the proportional hazards assumption was violated if this term was included, and the first events occurring were sepsis (169 restrictive group and 171 liberal group), wound infection (21 restrictive group and 26 liberal group), stroke (12 restrictive group and 12 liberal group), MI (one restrictive group and four liberal group), gut infarction (four restrictive group) and AKI (126 in restrictive group and 105 in liberal group).

The combinations of primary outcome components occurring are described in Table 22. Four hundred and eighty participants experienced one component, of which sepsis alone was the most common (240/480, 50%); 151 participants experienced two components (of which sepsis and AKI was the most common combination; 72/151, 47.7%) and 17 participants experienced three components (of which sepsis, wound infection and AKI was the most common combination; 11/17, 64.7%). More participants in the restrictive group had AKI alone than the liberal group (7.2% vs. 5.9%), but fewer had AKI and sepsis (3.1% vs. 4.1%).

Sensitivity analyses

Sensitivity analyses were pre-specified in the SAP, but not the study protocol; several of these were planned during data collection in response to knowledge about limitations of the study (e.g. non-adherence) or accruing data (e.g. inconsistency in data characterising renal function) but without any knowledge about how the sensitivity analyses would impact on the findings. The rationale and hypotheses for the sensitivity analyses have been explained previously (see Chapter 2, Sensitivity analyses).

The effect of non-adherence was assessed by estimating centre-specific treatment effects and ordering sites by rates of severe non-adherence (Figure 12). The hypothesis that estimates would tend towards the null with increasing non-adherence was not supported either visually, from the forest plot, or statistically in that a test for heterogeneity suggested no significant differences between sites (p = 0.65).

FIGURE 12.

Sensitivity analysis: primary outcome OR estimates by site, ranked by severe non-adherence rates. The grey vertical lines represent the overall treatment estimate (solid line) and 95% CI (dashed lines) of the primary outcome for the entire analysis cohort. The sizes of the point estimates reflect the centre sizes. Test for heterogeneity between sites; p = 0.65. Estimates have not been calculated for three sites with fewer than 20 participants with complete primary outcome data.

Excluding primary outcome events occurring in the first 24 hours after randomisation did not substantially alter the estimated treatment effect (Table 23), which did not support the hypothesis that the effect would tend away from the null. Excluding participants who received transfused red blood cells prior to randomisation caused the treatment effect estimate to increase (OR 1.23, 95% CI 0.97 to 1.54; p = 0.084), which was consistent with the hypothesis. Excluding AKI events without the relevant creatinine rise did not alter the treatment effect estimate; however, including additional AKI events not picked up via clinical assessment (anticipated to be milder events) caused the treatment effect estimate to increase (OR 1.20, 95% CI 1.00 to 1.44; p = 0.045) and the distribution of AKI events across AKI stages also became more pyramidal, consistent with adding in extra mild events. We had not hypothesised an increase in the treatment effect for this analysis but it was consistent with the observation that the small difference in the primary outcome frequency arose mainly from AKI events (see Table 23) and with the last planned sensitivity analysis, including only serious primary outcome events, which unexpectedly reduced the treatment effect estimate to unity (OR 0.99, 95% CI 0.77 to 1.27; p = 0.94).

Finally, the post-hoc analysis of severe in-hospital events (death, severe sepsis, ARDS, tracheostomy, low cardiac output, MI, AKI stage three, gut infarction and/or stroke) showed that this composite outcome occurred in 94/995 (9.5%) participants in the restrictive group and 90/993 (9.1%) in the liberal group (OR 1.05, 95% CI 0.77 to 1.43; p = 0.75).

Subgroup analyses

Subgroup analyses pre-specified in the study protocol are summarised in Figure 13. The subgroup analysis showing the largest difference between strata contrasted the treatment effect for participants with and without chronic obstructive pulmonary disease or asthma. The analysis suggested that the liberal transfusion strategy might be beneficial for participants with pulmonary comorbidity, although few participants had this comorbidity and the effect is not statistically significant. There was no other evidence of any subgroup effects.

FIGURE 13.

Subgroup analyses. The grey vertical lines represent the overall treatment estimate (solid line) and 95% CI (dashed lines) of the primary outcome for the entire analysis cohort. The sizes of the point estimates reflect the sizes of the subgroups. COPD, chronic obstructive pulmonary disease; LV, left ventricular; mod, moderate.

Primary analyses

Other clinical outcomes

There were significantly more deaths from any cause in the restrictive group (4.2%) than the liberal group (2.6%; HR 1.64, 95% CI 1.00 to 2.67; p = 0.045) (Table 24 and Figure 14). Causes of death and other SAEs that preceded death are given in Table 25. There are no clear causes of death contributing to the excess in the restrictive group; the common causes were cardiac disorders (21 participants), infections/infestations (13 participants) and general disorders and administration site conditions (10 participants). The primary outcome was experienced before death by 65% of participants. With respect to SAEs preceding death, 55% of the deaths in the restrictive group were preceded by an ischaemic SAE, whereas 58% of the deaths in the liberal group were preceded by an infectious SAE.

TABLE 24 - Other clinical outcomes

Post-randomisation stay only, which was (1) zero for 63 restrictive group participants and 61 liberal group participants, and (2) censored for 23 restrictive group participants and 15 liberal group participants. In addition, 37 restrictive group participants had more than one ICU/HDU admission (33 had two admissions and four had three admissions) and 32 liberal group participants (31 had two admissions and one had four admissions). A sensitivity analysis investigating the potential effect of informative censoring, assuming all censored patients had the maximum observed (i.e. not censored) duration of ICU/HDU stay, gave an estimated HR 0.94 (95% CI 0.86 to 1.03; p = 0.20).

Post-randomisation stay only, which was (1) zero for four restrictive group participants and two liberal group participants, and (2) censored for 25 restrictive group participants and 17 liberal group participants. A sensitivity analysis investigating the potential effect of informative censoring, assuming all censored patients had the maximum observed (i.e. not censored) duration of hospital stay, gave an estimated HR 0.98 (95% CI 0.90 to 1.07; p = 0.71).

Outcome	Randomised to restrictive threshold (N = 1000)	Randomised to liberal threshold (N = 1003)	Effect (95% CI)	p-value
All-cause mortality, n/N (%)	42/1000 (4.2)	26/1003 (2.6)	HR 1.64 (1.00 to 2.67)	0.045
Significant pulmonary morbidity, n/N (%)	127/979 (13.0)	116/982 (11.8)	OR 1.11 (0.85 to 1.45)	0.45
Initiation of non-invasive ventilation	88/989 (8.9)	77/984 (7.8)
Re-intubation/ventilation	50/975 (5.1)	53/973 (5.4)
Tracheostomy	30/988 (3.0)	32/988 (3.2)
Duration of ICU/HDU stay (hours),a median (IQR)	49.5 (21.9–99.7)	45.9 (20.1–94.8)	HR 0.97 (0.89 to 1.06)	0.53
Duration of hospital stay (days),b median (IQR)	7.0 (5.0–10.0)	7.0 (5.0–10.0)	HR 1.00 (0.92 to 1.10)	0.94

FIGURE 14.

Kaplan–Meier estimates of time from randomisation to all-cause mortality.

TABLE 25a - Further information on death: causes of death

CVA, cerebrovascular accident; GI, gastrointestinal.

Six participants have two primary causes of death listed: three participants in the restrictive group (participant 1, duodenal ulcer perforation and peritonitis; participant 2, multiorgan failure and sepsis; and participant 3, hypoxia and pulmonary oedema) and three participants in the liberal group (participant 1, endocarditis and ventriculocardiac shunt; participant 2, bronchopneumonia and empyema; and participant 3, bronchopneumonia and coagulopathy).

Causes of deatha	Randomised to restrictive threshold		Randomised to liberal threshold
	Events	Participants (n = 42)	Events	Participants (n = 26)
Blood and lymphatic system disorders	0	0	1	1
Coagulopathy	0	0	1	1
Cardiac disorders	10	10	11	11
Arrhythmia	4	4	1	1
Cardiac arrest	1	1	0	0
Cardiac failure	3	3	1	1
Cardiac failure acute	0	0	2	2
Cardiac failure congestive	1	1	4	4
Left ventricular failure	1	1	1	1
Left ventricular hypertrophy	0	0	1	1
Pericardial haemorrhage	0	0	1	1
GI disorders	6	6	1	1
Duodenal ulcer perforation	1	1	0	0
GI haemorrhage	0	0	1	1
Intestinal infarction	1	1	0	0
Intestinal ischaemia	3	3	0	0
Peritonitis	1	1	0	0
General disorders and administration site conditions	7	7	3	3
Multiorgan failure	7	7	3	3
Hepatobiliary disorders	1	1	0	0
Hepatic necrosis	1	1	0	0
Infections and infestations	8	8	5	5
Bronchopneumonia	1	1	2	2
Empyema	0	0	1	1
Endocarditis	1	1	1	1
Lower respiratory tract infection	1	1	0	0
Pneumonia	3	3	1	1
Sepsis	2	2	0	0
Neoplasms benign, malignant and unspecified	1	1	0	0
Brain cancer metastatic	1	1	0	0
Nervous system disorders	3	3	2	2
Cerebral haemorrhage	1	1	0	0
CVA	1	1	2	2
Haemorrhage intracranial	1	1	0	0
Renal and urinary disorders	1	1	0	0
Renal failure	1	1	0	0
Respiratory, thoracic and mediastinal disorders	4	4	0	0
ARDS	1	1	0	0
Hypoxia	1	1	0	0
Pulmonary oedema	2	2	0	0
Surgical and medical procedures	3	3	3	3
Cardiac operation	3	3	2	2
Ventriculocardiac shunt	0	0	1	1
Vascular disorders	1	1	3	3
Aortic aneurysm rupture	1	1	1	1
Haemorrhage	0	0	2	2

TABLE 25b - Further information on death: SAE preceding death

GI, gastrointestinal.

Excluding causes of death.

SAEs preceding deatha	Randomised to restrictive threshold		Randomised to liberal threshold
	Events, n	Participants (N = 42), n/N (%)	Events	Participants (N = 26), n/N (%)
Primary outcome		26/40 (65)		17/26 (65)
Sepsis		11/40 (28)		15/26 (58)
Wound infection		2/42 (5)		0/26 (0)
Permanent stroke		3/42 (7)		0/26 (0)
MI		0/42 (0)		1/26 (4)
Gut infarction		4/42 (10)		0/26 (0)
AKI		16/42 (38)		12/26 (46)
Transient ischaemic attack	0	0/42 (0)	1	1/26 (4)
GI complications	2	2/42 (5)	8	7/26 (27)
Post-operative haemorrhage	1	1/42 (2)	1	1/26 (4)
Cardiac tamponade	0	0/42 (0)	0	0/26 (0)
Pulmonary complications	23	14/41 (34)	29	10/26 (38)
Arrhythmias	15	12/42 (29)	9	7/26 (27)
Re-operation	12	11/42 (26)	4	4/26 (15)
Thromboembolic complications	1	1/42 (2)	0	0/26 (0)
Low cardiac output	6	6/42 (14)	10	6/26 (23)
Wound dehiscence	0	0/42 (0)	6	4/26 (15)
Other (unexpected event)	5	5/42 (12)	1	1/26 (4)
Cardiac arrest	1	1/42	0	0/26
Cardiac failure	3	3/42	0	0/26
Cardiac failure congestive	0	0/42	1	1/26
Compartment syndrome	1	1/42	0	0/26

In the restrictive group, 13.0% of participants experienced significant pulmonary morbidity compared with 11.8% participants in the liberal group (OR 1.11, 95% CI 0.85 to 1.45; p = 0.45). The median duration of post-randomisation ICU/HDU stay was 49.5 hours (IQR 21.9–99.7 hours) in the restrictive group and 45.9 hours (IQR 20.1–94.8 hours) in the liberal group (see Table 24). This difference was not statistically significant (HR 0.97, 95% CI 0.89 to 1.06; p = 0.53). Durations of total post-randomisation hospital stay were very similar in both groups [medians and IQRs for both groups were 7 days and IQR 5–10 days (HR 1.00, 95% CI 0.92 to 1.10; p = 0.94)].

Sensitivity analyses

The two sensitivity analyses outlined for the primary outcome that could be applied to the secondary outcome of mortality were performed on a post-hoc basis. The results are shown in Table 28. The treatment effects for both analyses, excluding deaths occurring in the first 24 hours after randomisation and excluding participants who had red blood cells transfused before randomisation, were shifted further away from unity, as was hypothesised.

Expected adverse events

Prior to hospital discharge (Table 30) there were 988 expected AEs in the restrictive group occurring in 49.6% of participants, and 938 in 48.4% of participants in the liberal group. Of these, 418 events in the restrictive group (23.2% of participants) and 405 events in the liberal group (21.5% of participants) were classified as serious. The most frequent events were arrhythmias. The numbers of pulmonary complications and arrhythmias were slightly larger in the restrictive group than the liberal group, which is consistent with Table 29.

Expected SAEs occurred less frequently after hospital discharge (Table 31), when there were 143 SAEs in 11.4% of participants in the restrictive group and 130 SAEs in 10.5% of participants in the liberal group. The most common events were pulmonary complications (5.0%). There were no clear differences between the groups although, again, rates of pulmonary complications and arrhythmias were slightly higher in the restrictive group than the liberal group (5.5% vs. 4.6% and 3.4% vs. 2.4%, respectively).

The classification of SAEs occurring at any time (either before or after discharge) suggests that the small differences in the frequencies of GI complications, pulmonary complications and arrhythmias identified between the groups were not due to any particular SAEs within each category (Table 32). Instead, all subcategories of SAEs appear to demonstrate differences between the groups in the same direction, which aggregate to the overall slight differences observed in Table 29.

Unexpected serious adverse events

There were 103 unexpected SAEs occurring in 8.8% of participants in the restrictive group and 113 events in 9.3% participants in the liberal group (Table 33). Classifying events according to the MedDRA dictionary suggests that cardiac disorders were the most frequent unexpected SAEs (3.0% of participants), in particular cardiac failure (30 events) and pericardial effusion (20 events). Other types of event were rare. There were more cases of anaemia in the restrictive group (six cases vs. one case) and more cases of cardiac failure in the liberal group (19 cases vs. 11 cases), although the numbers of both events were low. There were no other trends between the groups.