The Partners HealthCare electronic medical record (EMR) incorporates sociodemographic data, billing codes, laboratory results, problem lists, medications, vital signs and narrative notes from Massachusetts General Hospital and Brigham and Women's Hospital, as well as community and specialty hospitals which are part of the Partners HealthCare System in Boston (Massachusetts, USA). Altogether, these records comprise about 3.1 million unique patients.
Patients with the presence of at least one diagnosis of major depression determined by the presence of International Classification of Diseases, 9th edition codes (ICD-9 296.2×, 296.3×), in the billing data or outpatient medical record were selected from the EMR for inclusion in a data set (referred to as a data ‘mart’). The data mart consists of all electronic records (psychiatric and non-psychiatric) from 127 504 patients and can be utilised with the i2b2 server software (i2b2 v1.4).31
The present analysis included records from February 1990 to October 2009. The i2b2 system32
is a scalable computational framework, deployed at over 46 major academic health centres, for managing human health data, and the i2b2 Workbench facilitates analysis and visualisation of such data. The Partners Institutional Review Board approved all aspects of this study.
Subjects were classified into groups based upon documented antidepressant treatment from three sources: (1) medications prescribed to the patient via e-prescribing in the EMR, (2) medication documented but not prescribed by the documenting clinician and (3) medication dispensed by the inpatient pharmacy. The three primary groups were labelled ‘high’, ‘moderate’ and ‘low’ affinity for the serotonin transporter, based upon previous work.34
Classification of newer antidepressants, including duloxetine and escitalopram, was based upon Ki
—in the case of escitalopram, effective affinity is greater than citalopram because of the absence of the R enantiomer, which has demonstrated antagonistic effects.38–40
As reported affinities vary across publications,37
we prioritised comparative affinities reported by a single laboratory42
wherever possible and categories utilised in prior studies employing similar methodology to this one.35
Patients with multiple antidepressant prescriptions from different affinity groups were excluded from the analysis, as were those receiving tricyclic antidepressants or monoamine oxidase inhibitors, out of concern that these groups would not be well matched with those receiving newer treatments and to avoid the known cardiovascular risks associated with some older agents.
Antidepressant affinity groups by affinity for serotonin transporter*
For initial analysis, the ‘high’-affinity group was contrasted with the other two groups to yield two similarly sized cohorts for comparison, based on investigator consensus and recognising that other groupings might be equally reasonable. A further advantage of this distinction was that it assigned duloxetine and venlafaxine to different categories based on serotonin reuptake affinity, decreasing the likelihood of confounding by non-serotonergic (ie, noradrenergic) effects. The groups were first compared in terms of sociodemographic features, including age, gender and race and antidepressant use.
As electronic medical records data are not well suited to standard survival analytic approaches, we utilised an incident user cohort design to construct an exposure risk period for each patient. This method was proposed by Schneeweiss45
specifically for pharmacovigilance designs using electronic medical record and is conceptually related to prior approaches.46
This approach has previously demonstrated assay sensitivity in pharmacovigilance studies using similarly structured data.47
The exposure period begins on the date an antidepressant was prescribed and ends 30 days from prescription. If a second prescription is documented in the 30-day period, the period is extended another 30 days from this prescription. Patient analysis is censored as soon as their exposure risk period ends, that is, at the end of a continuous documented period of exposure. Outcomes occurring within any of the exposure periods are included in the analysis. Logistic regression was used to calculate crude RR, and RR was then adjusted for person-months of exposure and other potential confounding variables identified in the initial analysis of baseline characteristics. To examine the importance of this 30-day assumption, we also conducted a sensitivity analysis as a means of determining if the length of the exposure risk period had a meaningful impact on our analysis. We re-ran the analysis for all the outcomes with drug era windows of 60, 90 and 180 days. Adjusted RR and CIs are provided in table S3.
Three sets of analysis were performed. First, we examined associations between antidepressant group and vascular/bleeding events (primary outcomes) including gastrointestinal (GI) haemorrhage, myocardial infarction and stroke. Follow-up analysis examined each category of stroke (ischaemic or haemorrhagic) separately. For consistency with prior reports, we also conducted a secondary analysis on a subset of patients with more conservative criteria for selecting patients with major depressive disorder (MDD) (at lease two outpatient diagnoses or at least one inpatient diagnosis). Next, as a positive control or test for assay sensitivity, we examined association between aspirin exposure and GI bleed. Finally, as a negative control, we examined associations between antidepressant group and five outcomes selected by the clinical investigators (RHP, JWS, DVI and IK) based upon literature review as likely to be unrelated to serotonergic effects: acute liver failure, acute renal failure, asthma, breast cancer and hip fractures. Table S1 lists the ICD-9 codes used to identify these outcomes of interest.
All analyses utilised R 2.13.1 (The R Foundation for Statistical Computing).