Can we answer transfusion questions with retrospective data?
Over the past 2 decades, the increased focus on critical bleeding and trauma1, 2 and the emergence of protocol-based approaches to hemostatic resuscitation have suggested that platelet transfusion may be a key element in support and survival.3-5 Subgroup analyses of data from a randomized controlled trial of transfusion ratios in critically bleeding adult trauma patients suggest that early administration of platelets may improve outcomes.6 In other settings, even liver transplantation and cardiac surgery, the evidence is equivocal.7, 8 Therefore, given that the transfusion community is unlikely to attempt another large-scale randomized controlled trial anytime soon, we remain dependent on optimizing the thinking and tools brought to bear on retrospective observational studies.
In this issue of TRANSFUSION, Ning and colleagues describe their exploration of platelet transfusion in noncancer patients in intensive care units (ICUs). Specifically, this group examined the relationship between platelet transfusion and in-ICU and post-ICU mortality in a retrospective cohort of all adult patients admitted to ICUs in three academic medical centers in Hamilton, Ontario, from April 2008 to September 2015. Data for analysis were drawn from the Transfusion Registry for Utilization, Surveillance, and Tracking, a dedicated transfusion medicine database that merges medical, laboratory, and blood bank records data for all hospitalized patients in the McMaster consortium. Their final study cohort included 32,842 nononcology patients, 4927 (15.0%) of whom received at least one unit of platelets over the course of their ICU stay. Adjusted categorical (ever/never) and continuous (time-dependent) analyses suggested no association between platelet transfusions and increased ICU mortality (hazard ratio [HR], 0.78; 95% confidence interval [CI], 0.60-1.02; p = 0.06) or in-hospital mortality (HR, 0.89; 95% CI, 0.68-1.09; p = 0.41). The analysis suggested a possible association with improved overall hospital stay survival in general surgery patients. In this subgroup, platelet transfusion was associated with improved in-hospital survival when platelet exposures were evaluated dichotomously (HR, 0.71; 95% CI, 0.51-0.99; p = 0.04) and no harm when platelet exposures were evaluated continuously (HR, 0.93; 95% CI, 0.84-1.02; p = 0.10). Of note, general surgery patients comprised the smallest category of patients in the study cohort (23.4% of included patients were general surgery patients, compared with 48% medical patients and 28.6% cardiac surgery patients.) Also, general surgery patients received relatively few platelet transfusions (12.4% received platelets vs. 38.1% of cardiac surgery patients).
Considered in the wider context of retrospective clinical research in general and large databases of blood use and outcomes in particular, this study is as good as it gets: clear and well written; numbers of patients in the tens of thousands drawn from a unified national health care system into a multicenter, purpose-built registry; demographic, clinical, laboratory, and some injury severity scoring; appropriate analytic approaches thoughtfully applied and justified; and conclusions that are clear, useful, and not overstated. Critically, the authors distinguish between patients for whom platelets were presumably used therapeutically—that is, for active bleeding, especially postsurgically—and those for whom platelets were apparently used prophylactically, presumably in response to laboratory values or clinical concerns.
Two aspects of the statistical methods merit special attention. The first is the use of time-dependent stratification on platelet count in the multivariable Cox model. Although this prevents the authors from including platelets as a covariate, it respects the truth that platelet count and risk of transfusion vary closely over time, and also sidesteps any colinearity in the regression model with the baseline multiple organ dysfunction score (MODS, which is partly based on platelet counta). The second aspect worth attention is the avoidance of propensity score matching (PSM). The popularity of PSM has increased rapidly over the past 10 years as a method for approximating randomized cohorts, although recent research has cast serious doubt on its superiority as a matching method.9, 10 Regardless of these concerns, Ning et al. correctly point out that PSM would expose their analysis to immortal time bias, incorrectly considering some patients as exposed even though they had not been transfused with platelets yet.
That said, as in all registry-based work, these researchers had very limited control over the database itself. They could select fields to include but not add fields or augment them, and they were unable to review individual patient records to glean additional data of potential interest, leaving important concerns like the effects of pre-ICU component use—timing, products, volumes—unanswered. More importantly, roughly 10% of their study cohort were missing the MODS, their basis for controlling for illness severity. For prospective trials, current National Academies recommendations include setting minimum standards (and Data and Safety Monitoring Board compliance monitoring) for missing data and for approaches to imputation,11 but no such standards exist for retrospective studies, nor do Ning and colleagues describe any such approaches among their methods. In keeping with current practice recommendations, the authors do report that a “sensitivity analysis” did not contradict the findings of their primary analysis, despite the missing MODS scores. However, the assumptions and comparators on which this analysis is based are not given, so readers are left to decide how uncomfortable they are with 10% of a key variable—perhaps the key variable—being missing.
Blood use in or immediately after an unexpected critical bleeding episode is so tightly annealed to injury/illness severity as to preclude making valid causal associations between blood use and clinical outcomes unless the research model demonstrates an informed understanding of relevant measures of injury/illness severityb in critically bleeding patients. A range of literature suggests that blood component use in the first 4, 4 to 12, 12 to 24, and post-24 hours of acute trauma care targets significantly different patient groups. That is, as time to first blood products increases, patients are less injured, less likely to die, less likely to die of bleeding, and less sensitive to which blood product is hung first.12-14 Retrospective or prospective studies that conflate these groups or fail to account for the effects of survival biasc cannot be presumed to provide definitive answers to component use questions in critical bleeding situations. A limitation of the study by Ning and colleagues in understanding the role of platelet transfusion in the wider context of critical bleeding is that admission to the ICU is itself a strong marker of survival bias. By the time of ICU admission, patients have usually already survived the first few hours since the onset of life-threatening bleeding, a time during which the majority of patients who will die of bleeding do so12, 13 and where coagulation support—including platelets—can be lifesaving. This effect may be less pronounced in the cardiac surgery stratum, since essentially all of those patients are taken to the ICU postoperatively regardless of how sick they are.
Pre-ICU exposure and treatment effects are also likely different between groups. Most cardiac surgery patients will be admitted with some degree of post–cardiopulmonary bypass–induced platelet dysfunction that has been assessed and treated continuously, whereas the admission, platelet issues, and transfusion experience of those admitted after general surgery (including trauma) or with medical diagnoses may be quite different. Future retrospective work will also need to consider the effects of variations in platelet acquisition, storage, and pretransfusion manipulation and how those differences might play out in therapeutic vs. prophylactic platelet transfusion practice.
Overall, the study by Ning and colleagues regarding platelet transfusion and outcomes in ICU patients is well done, reasonably justified, and certainly a relief. But it should be understood not as an examination of the efficacy of platelet transfusion in treating or preventing critical bleeding but as an examination of the role of platelet transfusion as a potential modulator of the usual killers of ICU patients, that is, sepsis and multiorgan failure. Their study conclusions must be accepted as exactly what they are—reassurance about a common responsive practice, not a guidepost to the use of platelets in the resuscitation of life-threatening bleeding or the treatment of other emergent, high-risk, coagulopathic situations. That evidence awaits new work that understands and minimizes survival biases and optimizes control for injury severity, potentially critical variations in platelet product types, and the timing and order of product use from the onset of critical bleeding until it stops.
CONFLICT OF INTEREST
The authors have disclosed no conflicts of interest.
