Harms associated with single unit perioperative transfusion: retrospective population based analysis

Data source

The Premier Perspective database (Premier, Charlotte, NC) is an anonymized, fee supported, voluntary dataset created to support investigation of healthcare quality and utilization. It captures about 20% of all hospital discharges from acute care in the US. Data collected include standard hospital discharge file data and a date stamped log of all billed items, including drugs, laboratory and other diagnostic tests, and therapeutic services provided during the admission. The data are collected electronically and regularly audited by Premier. Procedure and comorbidity data are provided as ICD-9 (international classification of diseases, ninth revision) codes. Procedure codes are stratified into primary and secondary; discharge diagnosis codes are designated as primary or secondary and might have a modifier for “present on admission.” Our dataset included up to 25 diagnosis codes for each patient. A median of five (interquartile range three to eight) diagnoses were present on admission.

Patients

Patients in our analysis were admitted between 1 January 2009 and 31 March 2012, were aged ≥18, and were admitted for at least one night. We considered only the first episode of care for patients who had more than one admission during the study period. To attempt to exclude emergency cases, we did not include patients admitted from the emergency department or from hospital transfer. We also did not include patients with a primary diagnosis of intracranial injury or those undergoing obstetrical, cardiac, intracranial (including peritoneal shunt revision), or major vascular procedures, as those patients might have specific risks for stroke or myocardial infarction related to the procedure or diagnosis. We then excluded patients with a primary procedure code for a non-surgical procedure (such as intubation, thoracentesis, lumbar puncture) and patients with unknown sex. To avoid confounding by transfusions later in the admission, we also excluded patients who underwent transfusion on days two to seven after surgery.

Patient involvement

There was no patient involvement in this study.

Variables

Hospital level data directly collected by Premier include number of beds, urban or rural location, and teaching hospital status. Patient related values include year and month of admission, patient age, race/ethnicity, sex, length of stay and total hospital costs, insurance payor (whether the patient has health insurance provided through a private company, through the US government programs Medicare or Medicaid, or is uninsured), 30 day readmission, mortality (during index admission or on readmission), primary and secondary discharge diagnoses and primary procedure, and admission diagnosis if readmitted within 30 days. Diagnoses coded as present on admission were further categorized according to the methods of Elixhauser for obtaining pre-existing comorbid conditions from ICD-9 data.12

Details on transfusion of red blood cells were derived from billing data. Because some units were documented as partially transfused, the number of units of packed red blood cells was rounded up to the nearest integer, reflecting the number of units to which the patient was exposed. We included autologous, directed donor, and whole blood units, which in composite represented 3.3% of all units transfused.

Outcome definition

The primary outcome for this study was a composite of ischemic stroke, myocardial infarction, and ventricular tachycardia or fibrillation, which was included as a potential surrogate for coronary ischemia with or without myocardial infarction.13 Patients with a discharge diagnosis of stroke (ICD-9 433.x1, 434.x1, 977.02), initial episode of care for myocardial infarction (ICD-9 410.x1), or ventricular tachycardia/ventricular fibrillation (ICD-9 427.1 or 427.4x), or any of these diagnoses as a primary diagnosis for a subsequent readmission within 30 days, were considered to have had the primary outcome of interest. Strokes, myocardial infarctions, or episodes of ventricular tachycardia/ventricular fibrillation coded as present on admission during the initial admission were not included in the primary outcome. These ICD-9 codes, and therefore the outcome definition, do not include hemorrhagic stroke, subarachnoid hemorrhage, unspecified or subsequent episode of care for myocardial infarction, or atrial fibrillation.

Statistical analysis

Data management was conducted in SAS version 9.2 (SAS Institute, Cary, NC), and statistical analysis in Stata 13.1 (StatCorp LP, College Station, TX). Where not specified, P<0.05 was considered to indicate significance.

Primary logistic regression model

Variables were considered for entry into the model on the basis of face validity and a literature search to identify factors previously shown to be associated with perioperative stroke and myocardial infarction. A manual forward iterative technique was used to develop the multivariable logistic regression model, with the likelihood ratio test and area under the receiver operator characteristic (ROC) curve used to evaluate significant improvement of the model. We included in the model variables that achieved significant improvement in likelihood ratio at P≤0.025 and improved the area under the ROC curve by 0.005.

An initial non-hierarchical logistic regression model was first created based on stratification of the data into a population without cardiovascular risk factors (defined a priori as prior stroke or transient ischemic attack, coronary artery disease, congestive heart failure, atrial fibrillation, diabetes with chronic complications, renal failure, cardiac valvular disease, peripheral vascular disease, or chronic anticoagulant use). Significant associations of risk factors in the lower risk population were assumed to extend to the higher risk patients, and cardiovascular and other comorbidities were added into the model and selected for according to the above criteria. We generated a composite variable, incorporating history of congestive heart failure, atrial fibrillation, diabetes with chronic complications, and/or renal failure, because these comorbidities individually had low population prevalence (<5%) and grouping them as a composite variable compared with including them individually had minimal impact on area under the ROC curve, but in aggregate those factors produced an improvement consistent with inclusion according to the above criteria.

We then investigated potential interactions by sequential examination of population strata based on important model components (comorbidities or demographic factors). Interactions among the variables were investigated by using interaction terms in the logistic regression analysis and the margins and marginsplot commands in Stata, and significant interactions were retained.

Finally, the logistic regression model thus created was further refined by adjustment for random effects within hospitals with Stata’s meqrlogit command, which uses the QR decomposition of the variance components matrix to improve convergence. As the hierarchical component changed the odds ratios for the primary predictor by >10%, it was retained. The final hierarchical regression model had an area under the ROC curve of 0.833 and no evidence of collinearity with a maximum variance inflation factor of 1.37.

Subgroup analyses

The hierarchical model was also used to investigate the association between transfusion and stroke/myocardial infarction in surgical subgroups— total or partial colectomy, small bowel resection, hip or knee replacement or revision, laminectomy and/or fusion of the spine, and hysterectomy—on the basis of ICD-9 codes.

We used the same model in a subgroup restricted to patients who received no more than one unit of blood and in a subgroup in which we additionally excluded any patient who received transfusion on the day after surgery (that is, we included only patients who received transfusions on the day of surgery and those who had never received transfusion; the primary analysis already excluded patients who received transfusion on days two to seven after surgery). Finally, we analyzed outcomes individually (that is, stroke, myocardial infarction not including ventricular tachycardia/ventricular fibrillation, and ventricular tachycardia/ventricular fibrillation) to ensure odds ratios for all three outcomes were consistent and in the same direction, to confirm the appropriateness of the composite outcome analysis.

Propensity score analysis

We performed a propensity score analysis to complement the hierarchical logistic regression primary analysis. We used fixed logistic regression with clustering by hospital to compute a propensity score for any blood transfusion on the day of surgery or one day after based on all patient level factors in the primary model as well as chronic anticoagulant use, chronic lung disease, peripheral vascular disease, cancer, and inflammatory arthritis, again with random effects by facility. These factors were determined a priori on the basis of face validity and significant (P<0.025) univariate differences between groups who did and not receive transfusion. To improve predictive power of the propensity model, we additionally assembled a composite maximum surgical blood ordering schedule (MSBOS) using data from a published ICD-9-indexed partial MSBOS14 and supplemented with data from the MSBOS used at the University of California-San Francisco (UCSF).15 This schedule is indexed by a proprietary code used internally by our electronic medical record, which could not be linked to ICD-9 codes; accordingly, procedures with a frequency of >1000 in the dataset were manually matched to UCSF MSBOS values. A total of 157 592 patients undergoing 1512 distinct procedures were not matched to a MSBOS value. Area under the ROC curve for the propensity score model was 0.715. Then, 1-to-1 greedy matching without replacement was used to match all 41 421 patients who received a transfusion to 41 421 patients who did not, within a propensity score caliper distance of 0.01 or less in all cases. In the case of a tie, one patient was randomly selected from among the group of patients who did not receive a transfusion with well matched propensity scores.

We used this propensity score matched cohort in two ways. First, we performed univariate analysis comparing stroke/myocardial infarction rate between the transfused and non-transfused matched population. Then, because risk factors for stroke/myocardial infarction were non-significantly over-represented in the transfused cohort (leading to a possible overall bias toward stroke/myocardial infarction in the transfused group), we entered the matched cohort into the hierarchical logistic regression model developed above to adjust for the imbalanced covariates. Two propensity score sensitivity analyses are reported in table B in the appendix: one eliminating patients without a matched blood ordering schedule value (that is, inclusion only of patients with a known value), and one that does not use the blood ordering schedule. Results of those models were not substantially different to the model reported here.

Population attributable risk

The population attributable fraction (PAF) is a measurement of incremental increase in outcome rate experienced by patients with a certain risk factor compared with those without that risk factor. The adjusted fraction was calculated according to the methods of Greenland and Drescher16 with the Stata command punaf, which performs post-estimation calculations on models that produce conditional arithmetic means. The command is not compatible with interaction terms or hierarchical models, so we ran punaf on a simplified regression model omitting interaction terms and with clustering (not random effects) by hospital.

Estimating the impact of an unmeasured confounder

We used two methods to model the impact of a possible unmeasured covariate unequally distributed between the patients who did and did not undergo transfusion and the required strength of association with stroke/myocardial infarction required to invalidate the association between transfusion and this outcome. The first was according to methods described by Lin and colleagues,17 who use the algebraic relation between the modeled effect size in the primary model and the effect size and distribution of a hypothetical unmeasured confounder to model its impact. The second technique followed the methods of Sturmer and colleagues, which leverages propensity score calibration methods.18