Improving Reporting of Withdrawal of Life-Sustaining Treatment in Acute Brain Injury Trials: A Methodologic Imperative

For critically ill patients with acute brain injury (ABI), achieving survival with a neurologic outcome compatible with the patient’s wishes is a central priority. Accordingly, randomized clinical trials (RCTs) typically measure intervention effects using neurologic endpoints. Well-designed RCTs can support causal claims that an intervention reduces neurologic disability. For conclusions to be valid, trials must address familiar threats to validity, including protocol non-adherence, loss to follow-up, and inconsistent outcome ascertainment. As each of these issues can weaken causal claims if not carefully handled, trials must report these details transparently to support trustworthy conclusions, which is both a scientific and ethical responsibility (1).

A less familiar threat to causality in ABI trials is withdrawal of life-sustaining treatment (WLST). In critically ill patients with ABI, an estimated 60–70% of deaths are associated with a decision to withdraw life-sustaining treatment (2, 3), and nearly all patients experiencing WLST will die (4). WLST decisions are variable among sites and clinicians (3). Neuroprognostication is often central to these decisions, since WLST is frequently guided by the assumption that the anticipated neurologic outcome will not align with patients’ wishes. Yet while prognostic assessments incorporate markers of ABI severity, no clinical exam, test, or model provides perfect prediction of long-term outcome (5–7), and tools such as the Glasgow coma scale were designed for severity assessment on admission, not for prognostication (8). Some decisions are made in the very early phase of care when accurate neuroprognostication is even more difficult (9). WLST is also shaped by subjective factors such as clinicians’ heuristics, the framing of expected outcomes, institutional culture, and by resource availability or health insurance status (10). These influences do not disappear in RCTs, making it conceivable that neurologic outcome differences in some ABI trials could be influenced by level of care decisions.

In the context of WLST, bias may arise when patients’ “natural” outcome (i.e., their counterfactual outcome if interventions had not been withdrawn) is truncated by death (i.e., the consequence of most WLST decisions). Only the latter is observed in a trial. In some cases, a decision of WLST is made in patients who could have survived with a presumed favourable neurologic outcome (as defined in the trial), creating a difference between the counterfactual and observed outcome. Such “misclassification” events occur if prognosis is inappropriately pessimistic or if subjective factors guide decisions. On the other hand, they could also occur following a well-informed decision not entirely based on expected prognosis but on other patient-related factors. Evidence suggests that a non-trivial proportion of patients with ABI conditions may be misclassified (11, 12). Importantly, randomization does not protect against misclassifications because WLST is a post-randomization event.

A key implication of this phenomenon is that WLST may distort trial-reported treatment effects, potentially even reversing statistical or clinical interpretations (13). For example, in trials where clinicians are unblinded to treatment strategy, WLST may occur less frequently in the intervention group if clinicians believe the treatment is effective. Both the absolute number of deaths and the proportion of misclassified outcomes could therefore be reduced in the intervention group by mechanisms unrelated to the intervention. In this situation, if the intervention is not effective in improving neurologic outcome, it may appear effective as an artefact of lower incidence of WLST in that group. On the other hand, if an intervention is harmful (i.e., it truly worsens neurologic outcomes), WLST may reduce the perception of harm, because some of the poor outcomes related to the harmful treatment would then be balanced by fewer WLST-related deaths. The same mechanism suggests that, for effective interventions (i.e., treatment truly improves favourable neurologic outcome), the apparent benefit of the intervention could be amplified. Bias from WLST may also occur in trials where clinicians are blinded to the group assignment. Furthermore, decisions to WLST are often influenced by non-neurologic factors such as organ failure, previous co-morbidities, or altered quality of life. When these reasons bear an important weight in the decision to WLST, the risk of misclassification may be increased.

The myriad ways by which WLST can bias trial-reported treatment effects has prompted calls to uniformly disclose withdrawal and withholding characteristics in clinical trials (14). Cardiac arrest trials have started moving in this direction, based partly on the recognition that results of early hypothermia trials may have been influenced by WLST (15). However, reporting these data are just as important for trials in other ABI conditions, such as intracranial hemorrhage or traumatic brain injury, where WLST is common, occurs early, and is highly variable across institutions and countries (4).

Incomplete reporting of WLST in ABI trials makes it difficult to determine if the reported intervention effect is biased. This has serious implications for clinical practice, policy, and guidelines, particularly when the direction of treatment effect changes. We strongly encourage future ABI trials to adopt minimum reporting criteria in which the frequency, timing, and rationale for WLST decisions are reported (Table 1). Trialists should anticipate WLST as a critical variable, ensuring its systematic capture, and trial protocols should reinforce the importance of evidence-based, multi-modal, and transparent neuro-prognostication as part of standard clinical care. Journals could also accelerate progress by requiring WLST reporting as part of publication standards. Beyond descriptive reporting, trials should consider using analytic methods to quantify intervention effects accounting for WLST-related biases. Statistical approaches such as competing risks models, inverse probability censoring, and counterfactual simulation can be used to estimate what outcomes might have been, especially in the context of imbalanced WLST between groups (11, 12). These approaches cannot fully resolve uncertainty, but they can provide a credible upper and lower bound to treatment effects in a trial, and thus strengthen claims of causality.

Table 1:

Proposed minimum reporting criteria for reporting WLST and outcomes in ABI trials

Domain	Specific Elements to Report
Frequency	Number and proportion of WLST events in the overall trial, and number and proportion of deaths preceded by a WLST decision in each trial arm
Timing	Time from injury to WLST and time from randomization to WLST
Rationale	Reasons leading to the decision to WLST, incorporating assessment of physician prognosis and prior expressed patient wishes, where available
Sensitivity analyses	Exploratory analyses done in cases of WLST, such as imputation of outcomes based on observed characteristics, or inverse probability censoring, to provide a plausible range for the intervention effect

Although some might advocate refraining from WLST in the context of RCTs to reduce bias, such an approach would not be considered ethically or legally acceptable in most jurisdictions. Level of care decisions are an essential aspect of patient autonomy and inherently involve some degree of subjectivity. Precisely because these decisions cannot be standardized, and because WLST may be shaped by factors unrelated to objective neurologic prognosis, interpretation of treatment effects in RCTs becomes vulnerable to bias. The neurocritical care community must therefore recognize WLST as a key variable in trial design and analysis. Only by systematically capturing and transparently reporting the incidence, timing, and rationale of WLST can trials be appropriately appraised.