Population-based prospective cohort study
UK Biobank is a large-scale, long-term prospective study containing in-depth genetic and health information from half a million UK participants. Between 2006 and 2010, UK biobank enrolled 502,528 participants aged 37–73 years from 21 assessment centers across England, Wales, and Scotland. At recruitment, with their consent participants visited the closest assessment center to provide blood, urine, and saliva samples, as well as detailed information about sociodemographic, lifestyle and health-related factors, environment and medical history via touchscreen and face-to-face interviews. A range of physical measurements, including height, body weight, and blood pressure were taken. Follow-up assessments were conducted through linkages to routinely available national datasets. More details of UK Biobank design can be found elsewhere [23, 24]. The UK Biobank cohort has been approved by the North West Multi-center Research Ethics Committee, the England and Wales Patient Information Advisory Group, and the Scottish Community Health Index Advisory Group. All participants had provided written informed consent prior to data collection. In the present study, we excluded participants with stroke diagnosis prior to baseline (n=8750), and those who subsequently withdrew from the study (n=1299), leaving a total of 492,479 participants included in this analysis.
Assessment of PPI use
At baseline, regular use of PPIs was firstly assessed from participants using a touchscreen questionnaire and then confirmed during verbal interviews with a trained staff. In the touchscreen questionnaire, participants were asked “Do you regularly take any prescription medications?”. “Regular use” was defined as taking the medication in most days of the week for the last 4 weeks. If the participant selected “Yes” or “Unsure,” then they would be asked by the interviewer: “In the touch screen you said you are taking regular prescription medications. Can you now tell me what these are?” Information about PPI use was recorded in free text. The recorded type of PPIs included omeprazole, lansoprazole, pantoprazole, rabeprazole and esomeprazole. Information about doses and duration of PPIs was not collected. The detailed questions regarding PPI use could be found elsewhere .
Ascertainment of stroke
Participants were followed through linkage to the Health and Social Care Information Centre (in England and Wales) and the National Health Service Central Register (in Scotland). The primary outcome of the study was the incidence of stroke, which was linked to hospital admission and death registered using the International Classification of Diseases (ICD)-10 codes (I60, I61, I63, and I64). We classified stroke as ischemic stroke (I63, I64), intracerebral hemorrhage (I61), or subarachnoid hemorrhage (I60). Details of the methods used to identify stroke could be found on the UK Biobank website . At the time of analysis, complete follow-up was available up to 31 October 2017 for England and 31 October 2016 for Wales and Scotland.
Assessment of covariates
Covariate information was obtained at baseline. Sociodemographic factors (age, sex, ethnicity), lifestyle habits (smoking status, alcohol consumption, and dietary intake), family history of stroke, multivitamin use, and intake of mineral supplements were self-reported. Index of multiple deprivation, a composite measure of socioeconomic status, was provided directly from the UK Biobank. Physical activity was assessed using the International Physical Activity Questionnaire - Short Form (IPAQ-SF). Current concomitant comorbidities (hypertension, hypercholesterolemia, diabetes, prevalent cardiovascular disease [CVD] (including coronary artery disease, congestive heart failure, and peripheral vascular disease), atrial fibrillation, cancer, esophagitis/Barretts esophagus, GERD, and peptic ulcer), and medication use (aspirin, non-aspirin non-steroidal anti-inflammatory drugs [NSAIDs], acetaminophen, antihypertensive drugs (including angiotensin-converting enzyme inhibitors, angiotensin II receptor blockers, beta-blockers, calcium channel blockers, and thiazide diuretics), statin, metformin, histamine-2 receptor antagonists [H2RAs], antiplatelets, and clopidogrel) were assessed based on self-reported medical history, which were subsequently verified during face-to-face interview. Height and weight were measured by trained research staff and used to calculate body mass index (BMI). More details of these measures could be found elsewhere .
We calculated person-years from the recruitment date to the date of the first diagnosis of stroke, death, or the last date of follow-up, whichever happened first. We estimated the HRs of PPI use on stroke using Cox regression models taking age as the timescale. In the basic model, we stratified the analyses jointly by sex and age (37–54, 55–64, ≥65 years). In the multivariable-adjusted model 1, we adjusted for ethnicity, socioeconomic status, smoking status, alcohol consumption, physical activity, fruit and vegetable intake, BMI, multivitamin and mineral supplements intake, family history of stroke, history of hypertension, hypercholesterolemia, diabetes, CVD, atrial fibrillation, and cancer. We additionally adjusted for medications use (including aspirin, non-aspirin NSAIDs, acetaminophen, antihypertensive drugs, statin, and metformin) in the multivariable-adjusted model 2. To address the possible confounding effect of clinical indications for PPI use, we additionally adjusted for esophagitis/Barrett’s esophagus, GERD, peptic ulcer, H2RA use, and anticoagulant/antiplatelet use in the multivariable-adjusted model 3. Proportional hazards assumption was checked using Schoenfeld’s tests and no violation was shown. For covariates with selections of “do not know” and “prefer not to answer,” or with missing data, we included an “unknown/missing” value indicator. To present the association in a clinically useful way, we calculated the number needed to harm (NNH) and RD based on the method described by Altman D.G and Andersen P.K .
We also evaluated the baseline stroke risk of included participants using the Framingham Stroke Risk Score , based on which, we stratified the participants into subgroups of different risks. Then, we evaluated the relative effect (by HR) and absolute effect (by RD) of PPIs on stroke at each subgroup. We conducted additional stratified analyses according to sex, age, BMI, smoking status, alcohol consumption, physical activity, history of hypertension, hypercholesterolemia, diabetes, regular use of aspirin, history of GERD, and any clinical indications for PPI use.
We performed a number of sensitivity analyses to check the robustness of the primary results. First, we excluded participants who developed stroke or died during the first two years of follow-up to minimize reverse causality. Second, we excluded participants with cardiovascular disease or cancer to investigate the potential influence of the medical condition. Third, to evaluate potential bias from unobserved patient or physician characteristics (i.e., physicians may be more likely to prescribe PPIs to the patients with more severe underlying illness and also may be more likely to diagnose their patients with stroke in the appropriate clinical setting) [27, 28], we adjusted the number of self-reported operations, number of self-reported cancers, and number of self-reported non-cancer illnesses as surrogate indicators. Forth, we restricted the analyses to participants with no missing data on any covariates. Fifth, we calculated a propensity score for the likelihood of PPIs by multivariate logistic regression conditional on aforementioned baseline covariates. Then we applied inverse treatment probability weights based on the propensity scores, which creates a weighted pseudo cohort where treatment assignment is independent of measured confounders. To verify if potential biases could have modified the association between PPI use and stroke, we used falsification analyses for negative control outcomes (malignant melanoma cancer and transportation-related death) with the method described by Lipsitch M [29, 30]. We assumed that there should be no associations between PPI use and negative control outcomes. If these associations exist, the association between PPI use and stroke may be due to potential biases. We performed the analyses using SAS software, version 9.4 (SAS Institute, Cary, NC, USA).
We searched PubMed, EMBASE, and The Cochrane Central Register of Controlled Trials (CENTRAL, in The Cochrane Library) (from 1988 to 1 June 2020) for eligible studies, with no restriction in publication status and language. The search strategy was developed by an experienced group member (Jinqiu Yuan) and checked by two researchers from other teams (Zuyao Yang, The Chinese University of Hong Kong, China; Hongtao Wang, The Fourth Military Medical University, China) according to the PRESS 2015 Guideline Evidence-Based Checklist . The search strategy included terms for PPIs and a sensitive search strategy for randomized controlled trials, using the following combined keywords and MeSH terms: “proton pump inhibitors,” “omeprazole,” “esomeprazole,” “rabeprazole,” “pantoprazole,” and “randomized controlled trials” (see the complete search strategy for PubMed in Additional file 1: Table S1). We also searched the reference list of relevant review articles and included studies for additional eligible studies.
We included RCTs comparing PPIs with other interventions, placebo, or no treatment on stroke risk. Because the incidence of stroke is low in the population and small studies are unable to provide a reliable estimate of incidence, we only included trials that reported at least one case of stroke during follow-up, with a follow-up duration ≥ 6 months, and with a sample size ≥ 100. The outcome for meta-analysis was any stroke, included ischemic stroke, intracerebral hemorrhage, and subarachnoid hemorrhage. Study selection was undertaken by two authors (Man Yang and Qiangsheng He). We excluded trials about Helicobacter pylori eradication for the potential influence of antibiotics. Disagreements were resolved by discussion with a third reviewer (Jinqiu Yuan).
We initially imported all search citations into the reference management software and removed duplicated citations. We evaluated the eligibility of the remaining studies by examining the titles and abstracts. The full texts of potential eligible articles were retrieved to evaluate the eligibility. When two or more papers were published from a same study and the results were inconsistent, we only included the one with the largest sample size, most updated data, and the most relevant outcomes.
Two investigators (Qiangsheng He and Man Yang) extracted data and resolved disagreements by discussion. We extracted data with a pre-designed form for this study. The data extracted included study characteristics, methodological information, participant characteristics, intervention and control regimens, and outcomes. Missing outcome data were obtained by contacting authors and retrieving from clinical trial registries.
Assessment of risk of bias and quality of evidence
Two investigators (Qiangsheng He and Man Yang) evaluated the methodological quality of included studies using the Revised Cochrane Collaboration’s tool for assessing risk of bias (ROB 2) . The strength of evidence for primary estimates was evaluated using the Grading of Recommendations Assessment, Development and Evaluation system (GRADE) .
We undertook meta-analyses if included studies appeared appropriately similar in terms of patient population, intervention type, and outcome assessment. The summary effect size was measured as a risk ratio (RR), together with its 95% confidence interval (CI). We evaluated statistical heterogeneity with the Q-test and the I2 -index statistic. We carried out a meta-analysis with a fix-effect model (Mantel-Haenszel method). We evaluated publication bias with funnel plots and Egger’s test. We undertook sensitivity analyses to check the robustness of the primary result: (1) excluding studies with high risk of bias in one or more domains; (2) we excluded the COMPASS study which took up 94.3% weighting in the primary analysis. Meta-analyses were performed with Review Manager (Version 5.3. Copenhagen: Nordic Cochrane Centre, The Cochrane Collaboration, 2014).