Background: Aminian et al. recently reported an analysis of electronic health records (EHRs) showing that metabolic surgery is associated with a 39% reduction in the risk of major adverse cardiovascular events (MACE) compared with non-surgery, among obese patients with diabetes (JAMA 2019).
Objectives: We hypothesized that this previous finding is spurious due to differential information bias, since non-surgical controls do not undergo the same pre-diagnostic workup as patients undergoing surgery. To evaluate this, we replicated the original study and then modified its design so that the comparison groups had similar levels of baseline information. This was done by choosing a control group that consisted of patients undergoing an unrelated surgery (total hip/knee replacement) and using high-dimensional propensity score (HDPS) to match patients.
Methods: We conducted a cohort study using the Optum® de-identified Electronic Health Record dataset, an aggregation of inpatient and ambulatory records from 81 million patients between January 2007 through March 2018. Diabetes patients were selected as previously described. We performed an analysis that replicated the original study design, where metabolic surgery patients were compared against risk-set sampled non-surgery controls and matched by propensity score. We then conducted a subsequent analysis where the control group was changed to patients undergoing total hip/knee replacement and a high-dimensional propensity score (HDPS) model was used to match patients. The matched groups included 338 metabolic surgery and 547 hip/knee replacement surgery patients. Outcomes included 6-component MACE (mortality, coronary artery disease, cerebrovascular disease, heart failure, atrial fibrillation, or nephropathy).
Results: The results of Aminian et al. were replicated using the same analytical design (HR [95% CI] for MACE: 0.61 [0.55, 0.69] in the original study vs. 0.67 [0.61, 0.73] in the present study). When the comparison group was changed to patients undergoing hip/knee replacement and HDPS was used, there was no protective effect of metabolic surgery (HR [95% CI]: 0.99 [0.76, 1.30]) .
Conclusions: We demonstrated that information bias is the most likely explanation for the substantial cardiovascular benefit reported in the earlier study. Limitations in the present study include a small sample size in the matched comparison groups. Our findings highlight the importance of symmetric comparison groups when working with healthcare data without randomization.