Methods

Reproduced with permission from Hutchinson et al.6 © 2023 Massachusetts Medical Society.

Trial design

The RESCUE-ASDH was an international, multicentre, pragmatic, parallel-group randomised trial that compared the effectiveness of craniotomy versus DC for the management of adult head-injured patients undergoing evacuation of an ASDH. There was an additional observational cohort which included eligible patients enrolled into the study who could not be randomised after evacuation of haematoma due to significant swelling of the brain or who were enrolled following surgery.

Full ethical approval was obtained from each participating site. An internal pilot randomised trial was undertaken to confirm operational feasibility. Patients with an ASDH present in an incapacitated state, and due to its life-threatening nature, an operation is undertaken as soon as possible. Thus, a written agreement was obtained from the next of kin whenever feasible. If the next of kin was not available, the operation was undertaken in the patient’s best interest and agreement from the next of kin retrospectively obtained. An independent steering committee and an independent Data Monitoring and Ethics Committee reviewed the trial at regular intervals to assess conduct, progress and safety. The trial protocol was designed in a collaborative fashion by clinicians and researchers from the Academic Division of Neurosurgery and Anaesthesia at the University of Cambridge in conjunction with the SBNS and the European Brain Injury Consortium. The trial protocol was published previously and available with the full text of this article at NEJM.org.6 The trial investigators vouch for the accuracy and completeness of the data and the analyses and for the fidelity of this report to the trial protocol and the statistical analysis plan (SAP).

Participants

For enrolment in the study, patients had to be 16 years or older and have had an ASDH on a CT scan of the brain that required surgical evacuation with a large bone flap either by a craniotomy or by CT according to the opinion of the admitting neurosurgeon. Patients with additional lesions (such as intracerebral haematoma or contusions) could be included. Patients with bilateral ASDHs both requiring evacuation, previous enrolment in the trial, severe pre-existing disability or severe comorbidity which would lead to a poor outcome even if the patient made a full recovery from the head injury were excluded from trial participation. Trial sites were hospitals with acute neurosurgical services for patients with TBI.

Patients were enrolled into the trial from September 2014 through April 2019 in 40 centres in 11 countries. A total of 3566 patients were screened for eligibility, and 462 were randomised. Twelve patients were withdrawn due to lack of valid informed consent or withdrawal of consent. This resulted in a total of 228 in the craniotomy arm and 222 in the DC arm. There were 386 patients in the observational cohort: 141 received a craniotomy, and 245 received a DC.

Interventions

This trial compared two surgical procedures – craniotomy and DC. For the craniotomy arm, a large bone flap (recommended size ≥ 11 cm) anteroposterior diameter was raised and the ASDH evacuated. Other haematomas could be evacuated at the surgeon’s discretion. The bone was replaced and fixed to the surrounding skull with an appropriate fixation system. Hinge or floating flaps were not allowed. The size and location of the initial incision; method used to close the dura, galea and skin; use of drains and ICP monitors were left to the surgeon’s discretion. The DC arm followed the same procedure as the craniotomy arm, except the dura was left open or in a non-constricting manner and the bone flap was not replaced. Operations were undertaken by consultant neurosurgeons or trainees according to local practice. Medical management of these patients, pre, intra and postoperatively, were undertaken according to the centre’s standard practice for head-injured patients. The operating neurosurgeon assessed the suitability of randomisation in the operating theatre. Blocked randomisation was used, with a block size of four and allocation ratio of 1 : 1, and subjects were allocated randomly within each block. Allocation was stratified by country, geographical region, age, severity of injury and CT findings (see protocol and SAP for further details).6 Patients in the craniotomy arm could undergo a DC at a later stage if their condition deteriorated after their index procedure at the discretion of their treating physician. It was not possible to blind patients, relatives and treating doctors due to the nature of the operation (after a DC, a skull defect is noticeable until a cranioplasty is undertaken). However, the researchers assessing the outcome via follow-up questionnaires were blind to the operation undertaken.

Aims and objectives

Aim

The aim was to perform a multicentre, pragmatic, parallel-group randomised trial to compare the clinical and cost-effectiveness of DC versus craniotomy for the management of adult head-injured patients undergoing evacuation of an ASDH with an additional observational cohort for those patients’ undergoing surgery but who were unable to be randomised.

Primary objective

To compare the long-term clinical effectiveness of DC versus craniotomy (1-year follow-up period).

Secondary objectives

Compare the adverse events (AEs) and surgical complications between the two arms.

Undertake a detailed economic evaluation.

Outcomes

Primary end point

The Glasgow Outcome Scale-Extended (GOSE) at 12 months post injury. The use of the GOSE as a core global outcome measure is recommended by the interagency TBI Outcomes Workgroup and the International Mission for Prognosis and Analysis of Clinical Trials in TBI (IMPACT-TBI) group.

Secondary end points:

• GOSE at 6 months post injury
• HRQoL [EuroQol-5 Dimensions, five-level version (EQ-5D-5L)] at discharge from NSU, 6 and 12 months post injury
• Glasgow Coma Scale (GCS) on discharge from the ICU and from NSU
• LOS in ICU, neurosurgical and rehabilitation unit
• therapy intensity level (TIL) in the ICU
• discharge destination from NSU
• mortality
• serious AEs and surgical complications during index admission
• cranial surgery within 2 weeks after randomisation
• subsequent re-admissions to the NSU within the 12-month follow-up period
• hydrocephalus requiring shunt insertion within the 12-month follow-up period
• healthcare service utilisation over 12 months
• detailed economic evaluation.

Sample size

At the outset of the trial in 2014, the group of co-applicants deemed an 8% treatment effect as a clinically important difference. Subsequently, a formal sample size calculation was performed with the nQuery Advisor Version 7.0 (Statistical Solutions, Saugus, MA, USA) using a Mann–Whitney–Wilcoxon rank-sum test for ordered categories. We estimated that a sample of 990 patients would allow us to detect the equivalent of an absolute difference of 8% in the proportion of participants at 12 months after randomisation, with a favourable outcome [90% power and two-sided significance 0.05 (35% vs. 43%), 10% dropout rate] with ordinal analysis. Mortality is recorded as part of the GOSE (death – GOSE score 1), so it will not impact on loss to follow-up. In addition, a simulation study was undertaken to look at the impact on statistical power of covariate (age, GCS motor score and pupillary reaction) adjustment over and above the ordinal analysis. The simulation study confirmed that the planned sample size of 990 patients would give over 90% power to detect an 8% treatment effect.

However, it was clear by 2017 that the sample of 990 at the outset of the trial might not be appropriate given discussions in the neurosurgical community that evidence of a larger treatment effect is required to encourage surgeons to change their practice, especially when the intervention in question (i.e. craniectomy) will necessarily lead to a second operation (i.e. cranial reconstruction). A survey among an international pool of neurosurgeons who have expertise in neurotrauma and are interested in the study question was undertaken. The results showed that the mean number needed to treat (NNT) that would lead these surgeons to change their practice was seven. We believe that surgeons who do not have expertise or interest in neurotrauma are likely to suggest that an even lower NNT would be necessary in order to lead to a practice change. A NNT of seven is equivalent to a 14% treatment effect, and the sample size was re-estimated [using WinPepi version 11.65 (Joe Abramson, Hebrew University of Jerusalem, Jerusalem, Israel)] in 2018 using this treatment effect, giving a sample size of 440, allowing for a 10% dropout rate. The new sample size of 440 will have more than 80% power for a 14% treatment effect with binary analysis (corresponding to NNT of 7) and more than 90% power for a 14% treatment effect with ordinal analysis. The results of the survey and the implication for sample size adjustment were discussed at the HTA and Trial Steering Committee meetings, both unanimously agreed that reducing the sample size to 440 randomised patients was valid, based on these results.

If the target sample size of 440 is reached before 30 April 2019 (end of study recruitment date), we will continue recruiting patients until the end of the recruitment period.

Randomisation

Following enrolment in the study, suitability for randomisation was assessed in the operating room by the operating neurosurgeon. A secure web-based or telephone randomisation service was used for the randomisation of suitable patients. To randomise a patient, the following information was required: age, best pre-intubation GCS, pre-operative pupillary reactivity, CT findings.

Sequence generation

Blocked randomisation was used, with a block size of four and allocation ratio of 1 : 1. Subjects were allocated randomly within each block. Allocation was stratified by country/geographical region (UK/other European countries/North America/Australia/other), age group (< 40, 40–65, > 65 years), severity of injury (GCS 3–8, GCS 9–15) and CT findings (parenchymal injury, no parenchymal injury).

Implementation

The system provided an immediate allocation along with the patient identifier number for the trial. A confirmatory e-mail was sent to the e-mail addresses of the members of the study team at the site randomising the patient. The 24-hour randomisation service was also backed by 24-hour availability of trial investigators who could advise on patient eligibility.

Eligible non-randomised participants who had the operation (i.e. craniotomy or DC) deemed to be in their best interests by the operating surgeon were followed up in a similar manner to randomised patients in the observational arm of the study.

Blinding

Patients, relatives and treating doctors were unable to be blinded due to the nature of the intervention (after a craniectomy, a skull defect is noticeable until a cranioplasty is undertaken). The follow-up questionnaires were collected centrally (Cambridge), and outcome scores determined by two outcome adjudicators independently. The outcome adjudicators were blinded to patient allocation. Any disagreements were discussed in the outcome’s adjudication committee.

In the international sites, the same procedure was followed, with the difference that local staff were responsible for sending postal questionnaires to patients. If a telephone or face-to-face interview needed to be undertaken to complete the questionnaires, a standard operating procedure ensured strict separation between staff conducting these interviews and staff that are involved in the acute care of recruited patients. Questionnaires from international sites were returned to the Cambridge Clinical Trials Unit and outcome scores determined as described above.

Statistical methods

Primary outcome in this study will be measured using the GOSE. Conventionally, scales such as this are analysed by dichotomising the ordinal scale into a binary scale: ‘unfavourable’ versus ‘favourable’ and an odds ratio (OR) calculated. A significant number of patients will not have a realistic chance of crossing the threshold between ‘unfavourable’ and ‘favourable’ outcome and will not contribute data to the analysis. The crude OR is, therefore, not a meaningful effect measure for many patients. It discards much relevant information, reducing both the clinical relevance of the results and the statistical efficiency of the analysis. We, therefore, consider it more appropriate to quantify prognostic effects across the full range of the GOSE, and previous work by the IMPACT-TBI project has recommended that the primary analysis of the GOSE should use an ordinal approach, based on proportional odds methodology.7

Outcome analyses were performed in the modified intention-to-treat (ITT) population, which included all randomly assigned patients, except those who withdrew consent for participation in the trial and those lost to follow-up. Patients were retained in the group to which they were originally allocated, regardless of protocol adherence. The primary analysis was undertaken as an ordinal analysis using the proportional odds model, with the results presented as the estimated common OR with its corresponding 95% CI and p-value. The goodness of fit of the unadjusted proportional odds model was tested, and if rejected at 5% significance level, then that would mean that the model’s assumptions are not met.

The choice of ordinal logistic regression is based on the results of statistical research and trials, which showed that ordinal methods can increase statistical power substantially, equivalent to allowing a reduction of over 40% in the sample size without loss of statistical power. Ordinal logistic regression analysis is similar to logistic regression analysis, but it simultaneously estimates multiple ORs instead of just one. The number of ORs it estimates is equivalent to the number of ordered categories minus one. The ordinal regression model assumes that the OR for each potential split of the GOSE is constant no matter which cutpoint is taken, that is, it follows a proportional odds model; this assumption can be tested with a formal test of goodness of fit. The final estimated effect size is the common OR.

Further pre-specified secondary analyses were planned, including a fixed dichotomy analysis and a sliding dichotomy. The former compared the proportion of patients achieving a ‘favourable’ outcome (defined as upper severe disability or better on GOSE) between the two arms using the chi-squared test. The latter used a sliding dichotomy to define favourable outcome; if the randomisation GCS was between 3 and 8, a favourable GOSE outcome was defined as upper severe disability or better, but if the randomisation GCS was between 9 and 15, a favourable GOSE outcome was defined as lower moderate disability or better.

Analyses of secondary outcomes used proportional odds models for ordinal outcomes and chi-squared tests and logistic regression to estimate for binary outcomes. Chi-squared tests were used for categorical outcomes with categories combined, where necessary, due to low frequency, and Fisher’s exact tests were used for binary outcomes with low frequency in one category. Mann–Whitney U-tests were used for non-normally distributed continuous data and medians and interquartile ranges reported as the lower (Q1) and upper (Q3) quartiles. For LOS and survival data, Kaplan–Meier plots were produced and log-rank tests and Cox regression undertaken. Full details of the analysis are provided in the SAP (reference). Statistical analysis was undertaken using SPSS version 27 (IBM Corporation, Armonk, NY, USA).

Summary of any changes to the project protocol

The sample size was adjusted to 440, which had more than 80% power for a 14% treatment effect with binary analysis (corresponding to NNT of 7) and more than 90% power for a 14% treatment effect with ordinal analysis.