Data source and study population
We performed a retrospective observational study using a multicenter database in the United States (The eICU Collaborative Research Database). The database contains stratified random samples of patients (admitted between 2014 and 2015) from data repositories of 335 intensive care units (ICUs) from 208 hospitals located in the United States, amounting to 200,859 ICU admissions in total. Details of the database are described elsewhere (17). Eligible subjects were patients with sepsis at ICU admission. The definition of sepsis is in accordance with the third sepsis definition (1). Specifically, patients with documented or suspected infectious disease, along with the evidence of organ dysfunction (total Sequential Organ Failure Assessment [SOFA] score ≥ 2 points), were screened (1,18). Those who lacked data on platelet counts at ICU admission, or died, or were discharged from the hospital within 2 days after ICU admission were excluded from the analysis.
We used the REporting of studies Conducted using Observational Routinely collected health Data (RECORD) (19) statements for reporting this study.
Relative platelet reduction and Outcomes
Relative platelet reduction was calculated from the initial platelet count of day 1 (within 24 hours of ICU admission) and minimum values of day 2 (between 24-48 hours of ICU admission).
The primary outcome was in-hospital mortality. The secondary outcome was coagulopathy-related complications. The coagulopathy-related complications were defined as complications of new hemorrhagic events and thrombotic events after day 2 of ICU admission (Table S1 in additional file 1). If the active diagnosis of hemorrhagic and thrombotic events was recorded before and after day 2 in these patients, the patients were not considered to be experiencing new complications.
Covariates used for the primary model included patient-level variables such as age, sex, body mass index, ethnicity, comorbid conditions indicated by charlson comorbidity index, focus of infection, nosocomial onset of sepsis, severity score defined as the Acute Physiology and Chronic Health Evaluation (APACHE) IV score, number of ICU visits, and types of ICUs (20-22). Hospital-level variables including bed capacity, region of hospital location, teaching status of hospital were also included (23-25).
Continuous variables were expressed as medians with interquartile ranges (IQR) or means and standard deviations (SD), as appropriate. Categorical data were analyzed using χ2 tests to calculate unadjusted odds ratios.
For the primary analysis, multivariable mixed-effect logistic regression analysis was performed to estimate the mortality odds of each range of the relative platelet reduction (less than 20%, 20-40%, 40-60%, 60-80%, and more than 80%) and absolute platelet count on day1 (categorization was based on SOFA score). The hospital of admission was used as a random intercept to account for clustering by hospital (24,25). Subgroup analyses were also performed on those who did and did not experience thrombocytopenia, shock, and acute respiratory failure.
To further examine the association between platelet count trajectory and coagulopathy-related complications, multivariate logistic regression analysis was performed to estimate the odds of relative platelet reduction and absolute platelet counts on day 2 for coagulopathy-related complications. Since the incidence of coagulopathy-related complications were relatively rare, we did not adjust hospital-level variables in this model. The cutoff of ≧50% for the relative reduction and ≦100,000/μl for the absolute counts were used based on previous studies (13-15).
For handling missing data, we assumed that missing data were conditional, based on observed covariates (missing at random): multiple imputation was performed with multivariate imputation by the chained equations (MICE) package (26). For continuous, non-normal variables with upper and lower boundaries, we used predictive mean matching. The results of 10 imputed datasets were combined by averaging, and standard errors were adjusted to reflect both within- and between imputation variability.
For the sensitivity analyses, we repeated our analyses: 1) with the complete case set, 2) using models with additional covariates for adjustment, and 3) under alternative criteria of patient selection, to test the robustness of the primary analysis. The alternative model included the following additional covariates, which have the potential to be causally related to the outcome and/or platelet reduction: initial type of antibiotics, heparin administration, and renal replacement therapy. For the alternative cohort, we selected patients with an explicit diagnosis of sepsis (27). All statistical analyses were performed in R version 4.0.0.