Impact of multiple cardiovascular medications on mortality after an incidence of ischemic stroke or transient ischemic attack

A cohort study was conducted using The Health Improvement Network (THIN) database (now known as IQVIA Medical Research Data (IMRD)-UK database).

THIN is a primary care database that contains anonymised data from general practices across the UK. The database includes over 16 million patients from over 744 general practices. In 2013, the active patients in THIN represented approximately 6% of the UK population [14].

THIN includes information for each individual on demographics, diagnoses, prescriptions, referrals, laboratory tests, immunisations and the local area deprivation score (Townsend score) [15]. Primary care physicians and practice staff use a Read Code system to input and distinguish diagnoses, symptoms, investigations and lifestyle information in the electronic clinical notes. Prescription data are recorded via drug codes, and these can be identified by their generic name or by the British National Formulary (BNF) chapter [16]. THIN data have previously been used to study acute cardiovascular events [17].

This study included patients with their first diagnosis of ischemic stroke or TIA between January 2007 and December 2016. Patients who were aged 45 or above and who had been registered for at least 3 years in the THIN database before the first stroke event were included in this study. We excluded patients who had a history of myocardial infarction (MI) before the first stroke or TIA event, who had died of all causes or who had an occurrence of a further cardiovascular event within the first 90 days after the first event of stroke or TIA. Follow-up of the included patients commenced at the date of the incident stroke/TIA event and ended until the earliest of 31 December 2016, date of registered death and the date of leaving the general practice during the study period. For each patient, the follow-up was divided into contiguous periods of 1 year, each defined with specific entry and exit points.

Cardiovascular prescriptions were identified using drug codes in the THIN database. Each patient could contribute to several therapy categories, according to the cardiovascular medications issued at each entry point. Cardiovascular medications were identified based on all medications classified in the British National Formulary (BNF) Chapter 2 (cardiovascular system). Combination preparations were separated into their individual drug constituents.

We investigated the effect of combination therapy based on different numbers, classes and combination regimens on all-cause mortality. According to the numbers of cardiovascular medications (any medications identified based on BNF) prescribed in each 90-day exposure window, patients were stratified into groups of 0, 1, 2, 3, 4, 5 and ≥ 6 cardiovascular medications at each entry point. We then selected six evidence-based classes of cardiovascular medications commonly used for the secondary prevention of cardiovascular disease. The six classes of cardiovascular medications were antiplatelet agents (APAs), lipid-regulating medications (LRMs), angiotensin-converting enzyme inhibitors (ACEIs)/angiotensin receptor blockers (ARBs), beta-blockers (BBs), diuretics (DRs) and calcium channel blockers (CCBs) in stroke/TIA patients. Patients were stratified into groups of 0 (none of any cardiovascular medication) to 6 classes. Six classes of cardiovascular medications are APAs, LRMs, ACEIs/ARBs, CCBs, DRs and BBs exclusively. Patients who were on other class treatment were excluded from the study due to the complexity of the drug combination and few patients. Patients with one drug treatment or one class drug treatment were considered as the control group.

Patient demographics, clinical characteristics within 1 year prior to each entry point and prescriptions within 3 months prior to each entry point were extracted from the THIN database. Confounding variables included age, gender, smoking status (never smoked, former smoker), alcohol consumption (never drank, current drinker, former drinker), body mass index (BMI) (mean, normal, overweight, obese and underweight), blood pressure (BP) status (normal; stage 1, 2 and 3 hypertension; and hypotension), total cholesterol (TC) status (optimal, intermediate and high), Townsend scores, history of hypertension, hyperlipidaemia, arrhythmia, heart failure, peripheral vascular disease, percutaneous transluminal coronary intervention, diabetes, dementia, chronic obstructive pulmonary disease, asthma, liver disease, peptic ulcer disease, rheumatoid arthritis and chronic kidney disease. Previous use of cardiovascular medications and nonsteroidal anti-inflammatory medications (NSAIDs) were also included.

Data are summarised as mean (SD) for continuous variables and as frequencies (%) for categorical variables. Comparisons were performed using analysis of variance (ANOVA) for continuous variables and the chi-squared test for categorical variables. Multiple imputation was applied in addressing missing values for smoking status, alcohol consumption, BMI status, BP status, TC status and Townsend scores. We used multiple imputation by chained equations (MICE) in SAS version 9.4 to create 25 imputed datasets [18]. Rubin’s rules were applied to combine the results from analyses on each of the imputed datasets to produce estimates and confidence intervals [19].

We estimated the risk of mortality presented as hazard ratios (HRs) in relation to the number of medications, medication classes prescribed and different combinations using a marginal structural Cox proportional hazards model, as described by Hernán et al. [20]. This method aims to control for the effects of time-varying confounders and treatment switching. We estimated the parameters of our marginal structural model (MSM) by calculating a weight for each person-year interval and fitting a weighted pooled logistic regression model. Pooled logistic regression approximates the Cox model well when the risk of events is less than 10% per person-time interval [21]; herein, the maximum entry-specific risk of all-cause mortality was only 5.4%.

We used the inverse probability-of-treatment weight and the inverse probability-of-censoring weight to adjust for confounders at each entry point. In weight estimation, the numerator included the time-dependent intercept and the following baseline covariates: sex, baseline age, Townsend score, history of comorbidities and previous cardiovascular medications. The denominator included the time-dependent intercept, the baseline covariates and the following time-varying covariates: age at each entry point, most recently available smoking status, alcohol consumption, BMI status, BP status, TC status, comorbidities and previous occurrence of cardiovascular events (nonfatal MI, angina, stroke or TIA) 1 year prior to each entry point, and time-varying variables of previous cardiovascular medications and NSAIDs use 3 months prior to each entry point.

We conducted several sensitivity analyses: (1) using a 60-day screening period instead of a 90-day window, (2) dividing the 1-year follow-up time frame into intervals of 6 months, (3) including patients who had a history of MI before the first stroke or TIA event, (4) repeating the analyses in patients with completed characteristics data (complete-case analyses), (5) categorising missing data for each covariate as a separate group, (6) repeating the analyses separately for patients with TIA and patients with ischemic stroke and (7) an additional sensitivity analysis was conducted to assess the robustness of our findings to unmeasured confounding by computing the E value [22]. The E value is defined as “the minimum strength of association, on the risk ratio scale, that an unmeasured confounder must have with both the treatment and the outcome to fully explain away a specific treatment-outcome association, conditional on the measured covariates” [22]. All analyses were performed using SAS version 9.4.

The results of sensitivity analyses are provided in Additional file 1: Table S1-S2. Our primary results of the risk of mortality in patients with different numbers of cardiovascular medications and different numbers of classes of cardiovascular medications are similar to the results in the analysis using a 60-day exposure window. The analyses in patients with a history of MI, patients with competing risk characteristic data, when categorising missing data as a separate group and in separate analyses among patients with TIA only and patients with ischemic stroke only were consistent with the results of primary analyses. The results showed an even lower risk of mortality in patients with combination therapy when the follow-up duration was divided into 6-month intervals. The E values (risk ratios) for the three main analyses of all-cause mortality ranged from 1.74 to 4.57.

This study has several strengths. Firstly, this study was based on a large population-based primary care practice database. As such, it is likely to reflect the usual healthcare in the UK. Secondly, this study compared different numbers, classes and combinations of cardiovascular medications which comprehensively demonstrated the effect of combination therapy on long-term survival. Thirdly, when assessing the effect of different combinations, we defined exposure groups as patients who were exclusively using the selected cardiovascular medications of interest, and this was to remove potential effects of other cardiovascular medications which were not of interest on the outcome. In addition, we used MSMs to control for confounding due to both time-invariant and time-varying confounders that may lead to treatment switching or informative censoring. We demonstrated the robustness of our findings to unmeasured confounding using the E value estimate. Most HRs of all-cause mortality for known, strong risk factors of cardiovascular disease were below 1.74, the minimum E value in this study. For example, the HRs of mortality were 1.61 (95% CI 1.49–1.74) for current smokers, 1.27 (95% CI 1.19–1.36) for patients with diabetes and 1.14 (95% CI 1.07–1.20) for patients with hypertension. It is not likely that an unmeasured or unknown confounder would have a substantially larger effect on cardiovascular disease development or mortality than these known risk factors by having a relative risk exceeding 1.74. Finally, most compellingly, we used all-cause mortality as our outcome measure. Despite the influence of noncardiovascular mortality on the outcome, our study produced very clear results. Had we measured cause-specific cardiovascular mortality, we suspect that our findings would have been more pronounced.

This study has limitations. Firstly, the THIN database only provides records of prescriptions; therefore, our study was not able to determine if medications were actually dispensed, taken or used in line with the administration directions by patients. Secondly, because the THIN database does not capture data for hospital treatment, treatment in some care homes or nursing homes, and over the counter (OTC) medications (e.g. aspirin available OTC), the study was not able to address any medication usage not included in records from general practice. Thirdly, we had no information on the severity of stroke. Due to shorter life-expectancy, health interventions may be less cost-effective in patients with more severe cardiovascular conditions [34, 35]. In this case, patients with severe stroke may be more likely to be undertreated and thus more likely to die. However, we adopted measures to balance the heterogeneity between different exposure groups to some extent: (1) we excluded patients who had a history of MI before the first stroke event, (2) excluded patients who died or had a nonfatal cardiovascular event during the first 90 days after stroke or TIA and (3) we adjusted for risk factors of cardiovascular disease when estimating mortality hazard ratios. Fourthly, we only estimated the effect of cardiovascular medications by their major classification so our study cannot tell the effect of sub-classes of these cardiovascular medications on long-term outcomes. For instance, previous systematic reviews have suggested that dual antiplatelet therapy was more effective on short-term outcomes than monotherapy in stroke patients [9‐11], but our study did not compare the effect of dual-antiplatelet therapy and monotherapy on long-term mortality. Further research is required to explore this area. In addition, the clinical guidelines of pharmacotherapy for secondary prevention of stroke had no major changes over the period of 2007–2016 (refer to guidelines from AHA/ASA 2006 [36], 2010 [37] and 2014 [4]; National clinical guideline for stroke 2008 [38], 2012 [39] and 2016 [40]). There are some changes of recommendations on dosage and individual drugs. For example, in terms of lipid-lowering therapy in secondary prevention, the National Clinical Guideline 2008 recommended using statins according to a recommended cholesterol level. Guideline 2012 recommended high-intensity statin use such as atorvastatin 20–80 mg daily and Guideline 2016 recommended initiated using a statin with low acquisition cost such as simvastatin 40 mg daily. Our study only focused on the numbers and classes of CV drugs and did not address the dosage issue in the study due to the complexity of the research question and analysis. There may be some residual confounding impact on the mortality outcome in our study. But we would expect this impact is minimal. Future studies on drug dosage are encouraged.