A retrospective cohort study was conducted in the member population of Kaiser Permanente Southern California (KPSC), an integrated healthcare organization with over 7600 physicians, 15 medical centers, and 234 medical offices. KPSC provides comprehensive health care to over 4.6 million racially and socio-economically diverse members residing within seven counties of Southern California. Health care at KPSC is coordinated through region-wide electronic medical records (EMR) that capture detailed information about care provided to members at outpatient visits and during inpatient stays, as well as pharmacy, immunizations, imaging, and laboratory services received at KPSC-owned and contracting facilities. Our research database also includes administrative claims data for our members that capture any out-of-network clinical care and patient outcomes.
KPSC hospitals provide care to over 1 million ED patients per year (study sites ranging from ≈25,000 to 95,000 ED visits per year). Of these ED visits, approximately 80% are health plan members. All sites use the same troponin lab assay (Beckman Coulter Access AccuTnI+3) as well as a uniform 0.5 ng/ml threshold and a 0.04–0.5 ng/mL elevated risk cutoff.
The study was approved by the Institutional Review Board of KPSC.
Selection of Participants
We included all KPSC members aged 18 years or older with a visit for chest pain between 01/01/2015 and 12/01/2017 at 13 EDs operated by KPSC. To ensure complete comorbidity and outcomes capture, all included patients were required to have continuous health plan enrollment in the 12 months prior to and for at least 30 days post-discharge from their ED visit. ED encounters were included in the study if a valid troponin biomarker assay result was available for that encounter.
We excluded patients (Fig. 1) if they (1) had acute myocardial infarction identified using ICD9/10 codes, during the ED encounter; (2) had an initial troponin level greater than 0.5 ng/mL; (3) had invalid ED discharge status (e.g., against medical advice); (4) were transferred from another hospital; (5) died in the ED; (6) were in hospice status; (7) had a documented “do not resuscitate” order in the EMR.
Measurements and Outcomes
The primary outcome was the composite risk of 30-day acute myocardial infarction (see ICD9/10 codes in e-supplement) or all-cause death from the time of the initial ED visit. Death data were obtained from KPSC administrative records, EMR, and claims for out-of-network deaths. These data were supplemented with California state death files and Social Security Administration records for out-of-state deaths.
As our secondary outcome, we measured 30-day incidence of revascularization by percutaneous coronary intervention or coronary artery bypass grafting. We also measured 30-day incidence of acute myocardial infarction and death independently as secondary outcomes. Lastly, we defined major adverse cardiac event as the composite outcome of all-cause death, myocardial infarction, or revascularization within 30 days.
The 30-day time frame is consistent with ED acute coronary syndrome research guidelines as more extended time frames are unlikely to affect ED decision-making.14
The exposure was hospital admission for management of acute coronary syndrome, defined as either an inpatient or under observation status. We compared the effect of hospitalization disposition to discharge to home disposition.
Covariates included patient demographic information and clinical history. Age, sex, race, and insurance type were obtained from the health plan’s administrative records. Clinical data were obtained from the EMR. Comorbidities and cardiac risk factors were defined using laboratory values and diagnostic or procedure codes along with the Elixhauser index. The Elixhauser index15,16 is a well-validated comorbidity score, similar to the Charleson score, but more comprehensive. Body mass index (BMI) was measured from ED intake documentation or the most recently available visit, while smoking and family history of coronary artery disease/stroke were self-reported EMR fields. Those with a history of percutaneous coronary intervention or coronary artery bypass grafting were considered to have had prior coronary revascularization. Initial troponin level was dichotomized with a value below 0.04 ng/mL indicating a normal result and results between 0.04 and 0.49 ng/mL representing an elevated acute coronary syndrome risk. Lastly, using pharmacy prescription records, we identified patients on active antidiabetic, anticoagulants, anti-hyperlipidemia, and anti-hypertension treatment, in the 90 days prior to their ED encounter.
When using an observational study design, there remains a possibility of bias because some patients receive the treatment (or exposure) due to unrecorded factors strongly related to their prognosis. This bias creates a risk of confounding by indication. To mitigate this bias, we used the potential outcomes framework associated with the Rubin causal model (RCM) to evaluate the effect of hospitalization on death/acute myocardial infarction, revascularization, and major adverse cardiac event separately.17 We employed the generalized method of moments-based residual inclusion instrumental variables (IV) techniques to relax the restrictive RCM assumption of un-confoundedness.18,19 The residuals were based on a binary probit model that was used for the treatment choice (hospitalization vs. discharge to home) for the study cohort. GMM estimates a system of equations simultaneously and unlike multistep estimators, also provides correct standard errors for IV analysis in a single step.
We specified separate models for the binary outcomes associated with death, acute myocardial infarction, coronary revascularization, and major adverse cardiovascular events. All models were adjusted for age, sex, race, smoking, BMI, insurance type, self and family history of coronary artery disease, initial troponin, antidiabetic medication, anticoagulant medication, anti-hyperlipidemia medication, anti-hypertension medication, and Elixhauser comorbidities.
Based on prior research and previously validated methods20, we chose apriori to evaluate (1) the KPSC medical center’s historical practice pattern for hospitalization and (2) ED arrival time (categorized as 6 am–3 pm; 4 pm–11 pm, and 12 am–5 am), as two excluded instruments for the IV analysis, which we validated as part of our analysis.8 We postulated that patient arrival to ED during the late evening shift would make it more likely that the patient would be hospitalized as compared to those arriving early in the day. Each medical center’s practice pattern was calculated as the percent of suspected acute coronary syndrome patients who were hospitalized, in the 1 year prior to the ED date of each included cohort case with suspected acute coronary syndrome. The medical center’s practice pattern synthesizes consensus, experience and training of the ED professional staff, medical center’s protocol/policies, and available infrastructure for hospitalization. The calculation of the medical center’s practice pattern based on presenting patients’ ED encounter date made it dynamic and allowed capturing changes over time at the same medical center based on changes to any system or human capital factors (Supplementary Tables 1 and 2). Our final analysis was done using both of these instrumental variables.
We postulate that the time of ED arrival or population level medical center is unrelated to an individual patient’s death or myocardial infarction outcomes, except through the exposure. Therefore, we used these instrumental variables as a surrogate marker for the decision to hospitalize the patient or not, as a method to adjust for unmeasured patient or clinical factors that we did not expect to be affected based on these IVs. The IV specification testing presented in Supplemental Table 2 indicated that the two excluded instruments: (1) medical center practice pattern and (2) time of ED arrival were (a) strongly correlated to the treatment (i.e., hospital admission); (b) were not weak instruments; (c) satisfy the order as well as rank condition; (d) were not redundant and lastly were orthogonal to the outcome error and appropriately excluded from the outcome model since they only acted through the exposure of hospitalization.
We report the number need to treat (NNT) as the inverse of the adjusted absolute risk reduction (ARR) where: ARR = (absolute risk of outcomes for patients not hospitalized, i.e., controls) – (absolute risk of outcomes for patients hospitalized, i.e., intervention).
In the sensitivity analysis, we analyzed the data using doubly robust inverse probability of treatment weighted and regression adjusted (IPWRA) models assuming the un-confoundedness requirement was not violated. All hypothesis tests were two-sided with an a priori type I error set at 5%. The Stata/MP® version 15 software was used for data analysis (Stata Corp LLC, College Station, TX).