Mortality following myocardial infarction among HIV-infected persons: the Center for AIDS Research Network Of Integrated Clinical Systems (CNICS)

We used the Centers for AIDS Research Network of Integrated Clinical Systems (CNICS) cohort for all analyses. CNICS includes PWH in clinical care at 8 centers for HIV care across the USA. We included data from the 6 CNICS sites (Johns Hopkins University, University of Alabama at Birmingham, University of California-San Diego, University of California-San Francisco, University of North Carolina-Chapel Hill, and University of Washington) that completed adjudication of all detected incident MIs occurring between January 1, 1996, and December 31, 2014. Covariates in the analysis are drawn from the clinical data from all inpatient and outpatient encounters occurring for all HIV-infected persons at CNICS sites. Data in the CNICS repository include demographic data, HIV transmission risk factors, laboratory test results, such as HIV viral load, CD4+ T cell count, blood chemistries, cholesterol levels, and cardiac biomarkers; prescription medications, such as lipid-lowering therapies, antihypertensive medications, and antiretroviral medications; clinical diagnoses such as diabetes, hypertension, and smoking; patient-reported measures such as depression and alcohol use; and vital status including death dates. Each CNICS site obtained institutional review board approval for CNICS, and written informed consent was obtained from all CNICS participants. Demographic and clinical characteristics used for analyses were derived from the most recent visit prior to MI. There was no cap on the duration of time between the most recent visit at which a covariate was measured and MI with the exception of HIV viral load and CD4 T cell count, which were required within 3 years prior to MI for inclusion in this analysis. Otherwise, we carried forward the most recent pre-MI values over whatever time frame was appropriate. The median time between the pre-MI covariate values and MI was 2 months (mean 4 months).

Procedures for screening and adjudication of MIs have been described in detail [12, 22]. Possible MIs were identified by a screen for clinical MI diagnoses, coronary intervention documentation, or elevated cardiac biomarker levels. Each CNICS site also requested relevant medical records for any of their participants with reported or suspected MIs at outside hospitals. Sites then assembled de-identified packets of clinical notes, relevant cardiovascular and imaging studies, and laboratory tests which were reviewed independently by two expert physician adjudicators. Antiretroviral medication names were redacted to eliminate the potential for biasing reviewers based on perceived associations between specific antiretroviral medications and MIs. Two expert cardiologist adjudicators independently reviewed these data to determine whether CNICS participants had definite, probable, or no MI, and subsequently categorized events as T1MI or T2MI. Determination of MI presence and subtype included electrocardiographic criteria of evolving Q waves, ST elevations, and new left bundle branch block with specific algorithms used in the presence or absence of chest pain and cardiac biomarker elevations to determine whether MIs were definite, probable, or not present. Reviewers categorized MIs as T1MI or T2MI; definition of T2MI required identification of a clinical cause leading to the T2MI, such as an MI occurring during sepsis. A third expert cardiologist adjudicator was used to resolve discrepancies. Type 3 MIs were not included because these occur without available biomarkers, and there were < 10 MIs in the cohort that would have been categorized as type 4 or 5 MIs; thus, only type 1 and 2 MIs were included. Coronary interventions such as coronary artery bypass graft surgery were combined with T1MI. Only the first MI was included for analysis for CNICS participants with multiple MIs.

Deaths were ascertained for all CNICS participants at the local, state, and national level using multiple approaches including state death certificate data and national death indexes.

Follow-up started at the date of the MI and ended at the earliest of the date of death or date of censoring. The date of censoring was 18 months before the latest recorded death at each site in order to adequately ensure capture of deaths, given potential delays in death index data. All covariates were collected at least 7 days before the MI date to preclude MI-related treatment or laboratory changes to covariates. Due to this covariate collection date requirement, only MIs at least 7 days after the initial CNICS visit date were included. Regarding missing data, there were a total of 11 participants excluded because they did not have complete HIV viral load and CD4 T cell count data within 3 years prior to MI; the only other variable with missing data was body-mass index (BMI), which was 5.7% missing and for which missing data were handled with multiple imputation.

We used 3 approaches to identify potential predictors of death. First, Cox survival models were used to investigate the association of MI with covariates that were chosen a priori as clinically important. Second, to allow for variable selection into a more parsimonious statistical model and decide between potentially collinear predictors, Bayesian model averaging [26, 27] as applied to Cox models [28] was used to select potentially important covariates. We used a threshold of > 30% posterior probability of a predictor being included in the statistical model based on the goodness of fit of the family of possible models. Once the predictors were selected using Bayesian model averaging, Cox survival models with the selected covariates with high posterior probabilities of selection were used to estimate the associated hazard ratios for T1MI and T2MI. This was our primary analysis, as it allows a principled approach to selecting among highly correlated covariates. We also performed sensitivity analyses in which we repeated these analyses excluding persons who died within 30 days after MI in order to better isolate MI as an exposure rather than a near-concurrent outcome with death. Given the possibility that sepsis would precipitate many T2MIs and would affect post-MI mortality, we also performed a sensitivity analysis of mortality following T2MIs in which persons with and without sepsis were analyzed separately.