Tranexamic acid for acute gastrointestinal bleeding (the HALT-IT trial): statistical analysis plan for an international, randomised, double-blind, placebo-controlled trial

The HALT-IT trial is an international, randomised, double-blind (participants and trial staff), placebo-controlled trial to quantify the effects of TXA on morbidity and mortality in adults with significant upper or lower GI bleeding.

Pfizer Manufacturing, Marketing Authorisation number PL 00057/0952, manufactures the TXA. Torbay and South Devon NHS Foundation Trust, Manufacturing Authorisation number MIA (IMP) 13079, manufactures the placebo (sodium chloride 0.9%). Sharp Clinical Services (UK) Ltd., Manufacturing Authorisation number MIA (IMP) 10284, manufactures the study drug treatment packs containing either the active drug TXA or placebo. The Marketing Authorisation guarantees that the product is manufactured and released in accordance with the UK’s Good Manufacturing Practice (GMP) regulations. Ampoules and packaging are identical in appearance.

An independent statistician from Sealed Envelope Ltd. (UK) generates randomisation codes to be sent to Sharp Clinical Services UK Limited, a GMP-certified clinical trial supplies company that prepares trial treatment packs in accordance with the randomisation list. Sharp Clinical Services conduct the blinding process and first-stage Qualified Person (QP) release, which involves complete removal of the original manufacturer’s label and replacement with the clinical trial label bearing the randomisation number for use as the pack identification. Other pack-label text are identical for TXA and placebo treatments and in compliance with requirements for investigational medicinal products. Sharp Clinical Services UK are also responsible for maintaining the Product Specification File (PSF) until final database lock and unblinding of the trial data. Quality control checks to assure the blinding process are performed on a random samples of final QP released drug packs. High-performance liquid chromatography (HPLC) separation of known TXA is assessed against blinded samples to confirm which ampoule contains the placebo and active treatment. The tested samples are unblinded to assure accuracy of blinding.

The Trial Coordinating Centre (TCC) is responsible for assuring that all relevant approvals are available at the TCC before release of the trial treatment to a site. A separate Manual of Operating Procedures details the drug accountability system. The Investigator’s Brochure details labelling of the trial treatment and other processes for assuring adherence to GMP.

Eligible patients are randomised to receive either TXA or placebo as soon as possible and the study treatment started immediately. The next consecutively numbered treatment pack is taken from a box of eight packs. A fixed loading dosage of 1 g TXA or placebo (sodium chloride 0.9%) is administered, followed by a maintenance dose of 3 g TXA or placebo (sodium chloride 0.9%) infused over 24 h.

The trial was approved by the UK NRES Committee East of England (reference number 12/EE/0038), as well as national and local research ethics committees of participating countries outside of the UK.

Acute severe GI bleeding can be a frightening condition for the patient and the ensuing blood loss may have adverse impact on the patient’s mental and emotional state, impairing their decision-making ability. The consent procedures consider this together with the need to randomise and treat urgently. If the patient is fully competent, written consent is sought. If the patient’s capacity is impaired and a personal or professional representative is available, consent is sought from the representative. If neither are able to provide informed consent, consent is waived and the patient is informed about the trial as soon as it is possible.

The entry form (Additional file 1) is used to assess eligibility and collect baseline information. Once a patient has been randomised, the outcome in hospital is collected even if the trial treatment is interrupted or is not actually given. No extra tests are required but a short outcome form (Additional file 1) is completed from the medical records 28 days after randomisation or on discharge from the randomising hospital or on death (whichever occurs first). Any adverse events that become known to the investigator are reported up to 28 days after randomisation.

We originally specified all-cause mortality as the primary outcome because we believed that most deaths would be due to bleeding. However, as the trial was underway we observed that over half of all deaths were due to non-bleeding causes such as cancer and sepsis (see Fig. 1). Tranexamic acid reduces bleeding by inhibiting fibrinolysis. Based on this mechanism of action, we do not expect any substantial reduction in non-bleeding deaths. This hypothesis is supported by evidence from trials of TXA in trauma and postpartum haemorrhage [39, 40, 43]. As such, the treatment effect on all-cause mortality will be diluted by non-bleeding causes of death, reducing statistical power [43].

Death due to bleeding is the relevant endpoint for the HALT-IT trial because it has the potential to be reduced by the trial treatment. Fibrinolysis may play an important role in GI bleeding: gastric vein blood samples from patients with peptic ulcers contain high concentrations of plasmin and many patients with acute upper GI bleeding have elevated levels of fibrin degradation products (a biomarker for fibrinolysis) which is associated with worse outcomes [44‐46].

Cause of death is assigned by local investigators and a narrative of the events leading to death is reviewed by the principal investigator (who is blind to treatment allocation) and queried as necessary to verify cause of death. Due to the double-blind nature of the trial, the coding of the cause of death cannot be affected by the patient’s randomised group.

We also originally specified that the primary outcome would be measured up to 28 days after randomisation. However, patients receive TXA (or placebo) for their initial bleed but not for rebleeding episodes. Tranexamic acid has a half-life of 2–3 h so 99% will be eliminated within about 2 days of randomisation [47, 48]. We do not expect TXA to reduce deaths from a rebleeding episode for several weeks after the drug has been fully eliminated, therefore the primary outcome will consider early deaths due to bleeding only, defined as those that occur within 5 days of randomisation.

The rationale for refining the primary outcome from all-cause mortality to death due to bleeding was published in October 2018 [43]. The decision was supported by the Trial Steering Committee and was made prior to the end of the trial and prior to unblinding and so was not a data-dependent change.

The sample size calculation for the trial was based on the original primary outcome of all-cause mortality [42]. While the trial was underway, accumulating evidence from other large trials of TXA showed no apparent effect on non-bleeding causes of death [39, 40]. Because a considerable proportion of deaths in the HALT-IT trial are due to non-bleeding causes, the sample size was increased from 8000 to 12,000 to retain sufficient power for all-cause mortality. However, our assumptions were too generous – we assumed that 60% of deaths would be due to bleeding by the end of the trial rather than 40% (assuming a control group event rate of 10%, a study with 12,000 participants would have over 80% power to detect a 15% (RR = 0.6 × 0.75 + 0.4 × 1.0 = 0.85) reduction in all-cause mortality). Based on the refined primary outcome, assuming a cumulative incidence of death due to bleeding of 4%, a study with 12,000 patients will have about 85% power (two-sided alpha = 5%) to detect a clinically important 25% relative reduction in death due to bleeding from 4 to 3%. Loss to follow-up is expected to be less than 1% (it was 0.16% in the World Maternal Antifibrinolytic (WOMAN) trial) so was not taken into consideration when calculating the sample size. This power calculation is based on the primary analysis and refers to the unadjusted chi-squared test.

Patients with significant GI bleeding to whom the uncertainty principle applies are eligible. Specifically, a patient can be enrolled if the responsible clinician is substantially uncertain as to whether the trial treatment is appropriate for that particular patient. Significant bleeding is diagnosed clinically and implies a risk of bleeding to death. Patients with significant bleeding may include those with hypotension, tachycardia, signs of shock, or those needing urgent transfusion, endoscopy or surgery. Patients with a clear indication (e.g. traumatic haemorrhage) or contraindication (e.g. history of convulsions, thromboembolic disease) for TXA are excluded.

We will display the flow of study participants using a Consolidated Standards of Reporting Trials (CONSORT) Diagram (see Additional file 1: Figure S1). For each trial arm, we will present the total number randomised, the number with baseline data, the number lost to follow-up, the number who withdrew consent, and the number of participants with outcome data.

We collect data on the following baseline characteristics: age, biological sex, time from onset of GI bleeding symptoms to randomisation, suspected location of bleeding, clinical symptoms (e.g. haematemesis, melaena), suspected variceal bleeding, systolic blood pressure (SBP) , heart rate (HR), signs of shock, suspected active bleeding, major comorbidities, anticoagulation therapy and type of admission. We will present the distribution of baseline characteristics (n and %) in the treatment and placebo groups to check that randomisation was successful in producing similar groups (see Additional file 1: Table S1).

The main analyses will compare those allocated TXA with those allocated placebo on a modified intention-to-treat basis, excluding patients who received neither dose of the allocated trial treatment. We will present the results as effect estimates (relative risks) with a measure of precision (95% confidence intervals) (see Additional file 1: Table S2). Additionally, we will present results of the primary analysis adjusted for all baseline covariates. If baseline covariates are associated with the outcome, adjusting for any chance imbalances in baseline risk will increase statistical power. We will not present risk differences because they are not a generalisable measure of the treatment effect and are dependent on baseline risk. The effect of TXA will also be examined graphically using cumulative incidence curves presented with their associated hazard ratios and log-rank p values (see Additional file 1: Figure S2) [49]. The effects of TXA on death due to bleeding in the HALT-IT trial will be set in the context of other trials of TXA for acute severe haemorrhage (the CRASH-2 and WOMAN trials).

Death due to bleeding within 5 days of randomisation is the primary outcome. As outlined in the section ‘Change in primary outcome’ above, cause of death is assigned by local investigators who provide a narrative of the events leading to death. The cause of death narratives are reviewed by the principal investigator (who is blind to treatment allocation) and queried if more information is needed to confirm whether death is due to bleeding or another cause. Furthermore, due to double-blind nature of the trial, the coding of the cause of death cannot be affected by the patient’s randomised group. For more details, please see accompanying information in the section ‘Change of primary outcome’.

We will assess the effect of TXA on the following secondary outcomes. Unadjusted analyses will be presented in the main text and although we do not expect any baseline imbalances, to complement the unadjusted analyses and potentially increase statistical power (if covariates are associated with the outcome) we will present results of the analyses adjusted for all baseline covariates in an appendix.

We will conduct the following subgroup analyses for the primary outcome of death due to bleeding: time to treatment, location of bleeding, cause of bleeding and clinical Rockall score. We will fit interaction terms with randomised group in a Poisson regression model with robust error variance from the sandwich estimator [58]. Interaction tests (the Wald test) will be used to explore whether the effect of treatment (if any) differs across these subgroups. Results will be presented as unadjusted and adjusted effect estimates with a measure of precision (95% confidence intervals) and p value for the test for interaction (see Additional file 1: Table S4). Except for time to treatment, statistically significant heterogeneity between subgroups is required, as determined by the test for interaction p value, and not just statistical significance of a result in a specific subgroup [59].

Although treatment group is randomised within subgroups, the factors defining subgroups are not randomised. Several baseline characteristics are associated with the subgroup variables. For example, early treatment is correlated with bleed characteristics and patient characteristics (see Fig. 2), some of which confer a higher clinical Rockall score, suggesting that patients with more severe bleeding are treated earlier. Since these factors are also associated with mortality, they could potentially confound the interaction between time to treatment and the treatment effect.

If TXA is shown to be effective and the treatment effect varies by time to treatment, there is potential to intervene on time to treatment in order to increase the treatment effect. Although we cannot intervene on location of bleeding, cause of bleeding or clinical Rockall score, we are interested in ascertaining causal interaction of these factors with the treatment effect rather than simply assessing effect heterogeneity. As such, we will adjust all subgroup analyses for potential confounders [60]. Selection of potential confounders is based upon review of unblinded data within the trial to date in order to identify prognostic baseline characteristics that are associated with the subgroup variables. Potential confounders include age, time to treatment, SBP, HR, signs of shock, location of bleeding, suspected active bleeding, comorbid liver disease and suspected variceal bleeding. Signs of shock may be collinear with HR or blood pressure, and suspected variceal bleeding may be collinear with comorbid liver disease – if so, signs of shock and suspected variceal bleeding will not be included in the models. The final models remain to be determined because the outcome of interest is the treatment effect and the association between the potential confounders and the treatment effect cannot be assessed before unblinding.

Based on the data collected to date, we expect loss to follow-up to be minimal (i.e. less than 1% missing data on the primary outcome). Any missing values will be reported but not imputed.

We will conduct a survival analysis to explore the effect of TXA on rebleeding and death due to bleeding in more detail. In large trials of TXA for traumatic (CRASH-2) and postpartum haemorrhage (WOMAN), there were few late-bleeding-related events. The precise timing and duration of TXA’s antifibrinolytic effect remain to be determined. For example, it is unclear whether the treatment effect persists after the drug has been eliminated. Bleeding-related events occur later in acute GI bleeding, partly due to rebleeding, so the HALT-IT trial presents a unique opportunity to investigate this question.

We will report the median survival time and the cumulative incidence in the treatment and placebo groups, and model the treatment effect. Cox proportional hazards modelling assumes the hazards in the treatment and placebo groups are proportional over time. This assumption may be invalid if the antifibrinolytic effect of TXA declines over time as the drug is metabolised. We will formally assess this using the Royston-Palmer test for proportional hazards – a combined test with increased power when an early treatment effect is present [65]. If the treatment effect on death due to bleeding and rebleeding appears to change with time (non-proportional hazards), we will examine this in detail using various methods. We will estimate average cumulative hazard ratios for increasingly longer periods of follow-up. This method is preferred to period-specific hazard ratios, which can be susceptible to selection bias [66]. Nevertheless, we will also use Lexis expansion to calculate period-specific hazard ratios and test for interactions between treatment group and period. If we are able to identify the average duration of the treatment effect, we will examine whether this varies by baseline characteristics including time to treatment, bleeding severity, cause of bleeding and age. We will also assess how the treatment effect changes with time by including a time-by-treatment interaction term in the model. Residual methods will be used to test the assumption of linear time (first order trend) by plotting Martingale residuals against continuous covariates.

Death due to bleeding is a competing risk for non-bleeding causes of death and vice versa. Death is also a competing risk for rebleeding. We will estimate the treatment effect using a proportional cause-specific hazards model in which competing events are censored. The proportional cause-specific hazards model is preferred for aetiological research; however, both the cause-specific hazard and cumulative incidence can provide insights into a treatment’s effects [67, 68]. As such, a subdistribution hazards model and Gray’s test for comparing cumulative incidence functions will be presented as a supplementary analysis [69, 70]. Risk of rebleeding is highest immediately after the index bleed, death is a competing risk for rebleeding and some patients may experience more than one episode during the follow-up period. A survival analysis of the effect of TXA on rebleeding will take into account timing of events and competing risks.

If the trial demonstrates that TXA is an effective treatment for GI bleeding, we will conduct an economic evaluation to determine cost-effectiveness. Broadly speaking the methods will mirror those used by Li et al. who assessed the cost-effectiveness of TXA for the treatment of women with postpartum haemorrhage [71].

The analysis will compare TXA against clinical practice without TXA. A cost-utility analysis will be performed from a health services cost perspective with outcomes expressed as Quality-adjusted Life Years (QALYs). The analyses will be performed separately for a set of different countries, depending on where the majority of people have been recruited, but is likely to include at least the UK and Pakistan. A decision model will be used to extrapolate results from the trial into the longer term. Resource data, such as drugs and length of inpatient stay, are collected as part of the trial and will be analysed accordingly. Both deterministic and probabilistic sensitivity analysis will be undertaken. Results will also be presented by subgroups if considered appropriate.

Due to the large size of the HALT-IT trial, baseline characteristics should be well balanced between the treatment and placebo groups so that any difference in outcomes is due to the treatment. There is still a small possibility, however, that some imbalance in baseline risk may have arisen by chance. If prognostic factors are distributed differently across the treatment and placebo groups, this could bias the treatment effect. To investigate this hypothesis, we will conduct an analysis of the treatment effect on death due to bleeding adjusted for baseline risk. Patients will be stratified by risk deciles based on the predicted probability of death due to bleeding and a pooled effect estimate (relative risk) will be calculated using inverse variance weighting. This will provide an estimate of the treatment effect where both groups have equal baseline risks.

Centre- and country-level characteristics can influence patient outcomes. Differences in outcome may be related to resource availability or clinical practice. To explore between-country differences we will present a graph showing the number of patients and bleeding deaths by country and will use multivariable regression modelling to examine the treatment effect by country, including an interaction term between country and treatment. We will not adjust for clustering as we expect the effects of clustering to be small. Because we aim to understand any between-country differences in the treatment effect, we will adjust for potential confounders including age, SBP, HR, comorbidities, location of bleeding, suspected variceal bleeding, suspected active bleeding and time to treatment. A comparison between low-, middle- and high-income countries will be included using the World Bank country groupings by income. We do not expect the effect of TXA on the risk of death due to bleeding to vary by country, even though the absolute risk will vary due to between-country differences in patient populations. Countries recruiting fewer than 100 patients will be omitted from the analysis as necessary.

Between-centre differences in outcome may also influence the estimation of the treatment effect. We will first use a mixed-effects regression model using restricted maximum likelihood estimation to examine whether there are differences in death due to bleeding between centres. Results will be presented in the form of a forest plot. Prognostic patient characteristics (age, SBP, HR, comorbidities, location of bleeding, suspected variceal bleeding, suspected active bleeding), treatment group and time to treatment will be adjusted for. To take into account country-level effects we will also consider between-centre differences in outcome adjusted for country. We will then use mixed-effects regression to estimate the treatment effect before and after accounting for between-centre differences, assuming a constant treatment effect across centres. To assess whether the treatment effect differs by centre, we will fit a model with an interaction term between centre and treatment.