Impact of PPIs on GI Bleeding in Post-MI Patients on NSAIDs
Impact of PPIs on GI Bleeding in Post-MI Patients on NSAIDs
Diagnostic data came from the Danish National Patient Registry, which using ICD-10 (international classification of diseases, 10th revision) to classify hospital admissions (supplementary table 1). Each hospital admission is registered with one main discharge diagnosis and one or more supplementary diagnoses if appropriate. Information on vital status (dead or alive) came from the civil registration system through Statistics Denmark. We obtained primary, secondary, and contributing causes of death recorded by a physician from the National Causes of Death Registry. The National Prescription Registry provided information on the date of dispensing, quantity dispensed, strength, and formulation of all drugs dispensed from Danish pharmacies and classified according to the Anatomical Therapeutic Chemical (ATC) system (supplementary table 2). The partial reimbursement of drug expenses by the Danish healthcare system requires all pharmacies to register each drug dispensed in the National Prescription Registry, ensuring complete registration. In Denmark, every resident has a permanent unique civil registration number that enables linkage across administrative registries.
We identified all patients aged 30 years and over in the National Patient Registry between 1997 and 2011 who had a primary diagnosis of acute myocardial infarction (ICD-10 code I21 to I22), received antithrombotics, and survived at least 30 days from discharge (date of inclusion). The 30 day restriction defined a quarantine period. The follow-up period started after the 30 day quarantine period to minimise risk of immortal time bias. To avoid selection bias, we used a new user design, excluding patients who collected a prescription for an NSAID during the quarantine period (n=5711). Patients were followed until one of the following events (whichever came first): event of interest, emigration, death, or end of study period (31 December 2011). The diagnosis of myocardial infarction has been validated with a specificity exceeding 90%.
We used claimed prescriptions of aspirin or clopidogrel to characterise patients as receiving either antithrombotic monotherapy (aspirin or clopidogrel) or dual therapy. We excluded post-myocardial infarction patients who received no antithrombotic (n=15 353) or an oral anticoagulant only (n=1440). We identified all claimed prescriptions for NSAIDs (ATC M01A, excluding glucosamine (M01AX05)). We categorised rofecoxib and celecoxib as cyclo-oxygenase-2 selective (COX 2) inhibitors; ibuprofen, diclofenac, and naproxen as non-selective NSAIDs; and all other NSAIDs as "other" NSAIDs. The only NSAID available in Denmark without prescription is ibuprofen (since 2001). We categorised PPI treatment in one group (ATC A02BC); we also examined five individual PPIs—omeprazole, lansoprazole, pantoprazole, esmoprazole, and rabeprazole.
We calculated exposure periods for NSAIDs and PPIs for each patient by estimating a daily dose after comparing the accumulated dose and the elapsed time from consecutive prescriptions for the drug under investigation. We determined ongoing exposure by dividing the number of tablets/capsules dispensed by the estimated average dosage. If only one prescription was registered for an individual, we used a standard dosage, defined as the minimal recommended dosage, to estimate the daily dose. We used information on increasing or decreasing dosage only to continuously assess whether tablets were available. We defined exposure as having occurred when patients had drug available and discontinuation as when they had no more drug available. Methods for determining dose and treatment duration have been described previously. For most patients, treatment regimens changed during the study period, so we treated NSAID and PPI use in the analysis as time varying exposures—that is, patients changed exposure group according to claimed prescriptions. Each patient's exposure group at inclusion defined baseline treatment, shown in the Table with the covariante distributions (supplementary figure).
We identified comorbidities from previous diagnoses and at discharge from the index myocardial infarction, as specified in the Ontario acute myocardial infarction mortality prediction rule, and potential risk factors for bleeding (previous bleeding, alcohol consumption, liver disease, and ulcers). The Ontario acute myocardial infarction mortality prediction rule is a logistic regression model that predicts 30 day and one year mortality by using 11 variables determinable from hospital discharge databases (age, sex, shock, diabetes with complications, congestive heart failure, cancer, cerebrovascular disease, pulmonary oedema, acute renal failure, chronic renal failure, cardiac dysrhythmia). In our analyses, we incorporated each variable as a covariate and permitted diagnoses up to one year previously (supplementary table 1). To account for accumulation of risk factors during follow-up, we did an analysis including all variables in the main analysis as time dependent (exempt for inclusion year and percutaneous coronary intervention status (considered as related to myocardial infarction inclusion criteria) and previous peptic ulcer disease (considered an intermediate variable of the primary outcome during follow-up)). During the entire span of the database, patients could change status on the date of exposure.
We defined the primary outcome of gastrointestinal bleeding as hospital admission for or death from a bleeding gastrointestinal ulcer, haematemesis, melena, or unspecified gastrointestinal bleeding from the National Causes of Death Register and National Patient Register. Occurrence and type of bleeding as recorded in hospital databases have shown a positive predictive value of 89–99%.
We calculated crude incidence rates as number of events per 100 person years according to the different treatment regimens. We estimated the effects of PPI treatment on gastrointestinal bleeding with adjusted Cox proportional hazards models in terms of hazard ratios and 95% confidence intervals for gastrointestinal bleeding with the drug exposure continuously updating—that is, as time varying exposure allocated according to treatment regimen. We considered patients to be at risk only when exposed to the drug (during active treatment). Each patient could have multiple treatment groups throughout follow-up. We calculated risk time (person years) only for the active treatment period. The timescale in the Cox model was days passed since inclusion. We adjusted all models for age, sex, year of index hospital admission, concomitant drugs, comorbidity, and percutaneous coronary intervention status. We did additional analyses to assess any association between PPI use and individual antithrombotic regimens and NSAIDs.
We did nine sensitivity analyses. (1) To take account of any effect of over the counter NSAID use, we ended follow-up at 2001. (2) To take account of dabigatran or ticagrelor use (released in Denmark, August 2011), we ended follow-up in December 2010. (3) We examined cardiovascular death as a solo endpoint. (4) We stratified the cohort at 65 years to take account of guidelines recommending PPIs for people over 65 taking antithrombotic treatment. (5) We controlled for the variables included in the HAS-BLED score. (6) We did an analysis including all covariates as time dependent. (7) We stratified the population in two groups: high (previous bleeding) and low (no previous bleeding) risk of gastrointestinal bleeding. (8) Although the indication for NSAID use was not systematically available, we were able to do an analysis of patients with rheumatoid arthritis. (9) We examined the effect of duration of NSAID treatment (0–14 days and >14 days).
We confirmed the validity of the proportional hazard assumption, linearity of continuous variables, and lack of interaction and chose a significance level of 0.05. We used SAS 9.2 and Stata 11.0 for all statistical calculations.
No patients were involved in setting the research question or the outcome measures; nor were they involved in the design and implementation of the study. There are no plans to involve patients in dissemination.
Methods
Data Sources
Diagnostic data came from the Danish National Patient Registry, which using ICD-10 (international classification of diseases, 10th revision) to classify hospital admissions (supplementary table 1). Each hospital admission is registered with one main discharge diagnosis and one or more supplementary diagnoses if appropriate. Information on vital status (dead or alive) came from the civil registration system through Statistics Denmark. We obtained primary, secondary, and contributing causes of death recorded by a physician from the National Causes of Death Registry. The National Prescription Registry provided information on the date of dispensing, quantity dispensed, strength, and formulation of all drugs dispensed from Danish pharmacies and classified according to the Anatomical Therapeutic Chemical (ATC) system (supplementary table 2). The partial reimbursement of drug expenses by the Danish healthcare system requires all pharmacies to register each drug dispensed in the National Prescription Registry, ensuring complete registration. In Denmark, every resident has a permanent unique civil registration number that enables linkage across administrative registries.
Study Population and Follow-up
We identified all patients aged 30 years and over in the National Patient Registry between 1997 and 2011 who had a primary diagnosis of acute myocardial infarction (ICD-10 code I21 to I22), received antithrombotics, and survived at least 30 days from discharge (date of inclusion). The 30 day restriction defined a quarantine period. The follow-up period started after the 30 day quarantine period to minimise risk of immortal time bias. To avoid selection bias, we used a new user design, excluding patients who collected a prescription for an NSAID during the quarantine period (n=5711). Patients were followed until one of the following events (whichever came first): event of interest, emigration, death, or end of study period (31 December 2011). The diagnosis of myocardial infarction has been validated with a specificity exceeding 90%.
Antithrombotic, PPI, and NSAID Treatment
We used claimed prescriptions of aspirin or clopidogrel to characterise patients as receiving either antithrombotic monotherapy (aspirin or clopidogrel) or dual therapy. We excluded post-myocardial infarction patients who received no antithrombotic (n=15 353) or an oral anticoagulant only (n=1440). We identified all claimed prescriptions for NSAIDs (ATC M01A, excluding glucosamine (M01AX05)). We categorised rofecoxib and celecoxib as cyclo-oxygenase-2 selective (COX 2) inhibitors; ibuprofen, diclofenac, and naproxen as non-selective NSAIDs; and all other NSAIDs as "other" NSAIDs. The only NSAID available in Denmark without prescription is ibuprofen (since 2001). We categorised PPI treatment in one group (ATC A02BC); we also examined five individual PPIs—omeprazole, lansoprazole, pantoprazole, esmoprazole, and rabeprazole.
We calculated exposure periods for NSAIDs and PPIs for each patient by estimating a daily dose after comparing the accumulated dose and the elapsed time from consecutive prescriptions for the drug under investigation. We determined ongoing exposure by dividing the number of tablets/capsules dispensed by the estimated average dosage. If only one prescription was registered for an individual, we used a standard dosage, defined as the minimal recommended dosage, to estimate the daily dose. We used information on increasing or decreasing dosage only to continuously assess whether tablets were available. We defined exposure as having occurred when patients had drug available and discontinuation as when they had no more drug available. Methods for determining dose and treatment duration have been described previously. For most patients, treatment regimens changed during the study period, so we treated NSAID and PPI use in the analysis as time varying exposures—that is, patients changed exposure group according to claimed prescriptions. Each patient's exposure group at inclusion defined baseline treatment, shown in the Table with the covariante distributions (supplementary figure).
Comorbidity
We identified comorbidities from previous diagnoses and at discharge from the index myocardial infarction, as specified in the Ontario acute myocardial infarction mortality prediction rule, and potential risk factors for bleeding (previous bleeding, alcohol consumption, liver disease, and ulcers). The Ontario acute myocardial infarction mortality prediction rule is a logistic regression model that predicts 30 day and one year mortality by using 11 variables determinable from hospital discharge databases (age, sex, shock, diabetes with complications, congestive heart failure, cancer, cerebrovascular disease, pulmonary oedema, acute renal failure, chronic renal failure, cardiac dysrhythmia). In our analyses, we incorporated each variable as a covariate and permitted diagnoses up to one year previously (supplementary table 1). To account for accumulation of risk factors during follow-up, we did an analysis including all variables in the main analysis as time dependent (exempt for inclusion year and percutaneous coronary intervention status (considered as related to myocardial infarction inclusion criteria) and previous peptic ulcer disease (considered an intermediate variable of the primary outcome during follow-up)). During the entire span of the database, patients could change status on the date of exposure.
Outcome
We defined the primary outcome of gastrointestinal bleeding as hospital admission for or death from a bleeding gastrointestinal ulcer, haematemesis, melena, or unspecified gastrointestinal bleeding from the National Causes of Death Register and National Patient Register. Occurrence and type of bleeding as recorded in hospital databases have shown a positive predictive value of 89–99%.
Statistical Methods
We calculated crude incidence rates as number of events per 100 person years according to the different treatment regimens. We estimated the effects of PPI treatment on gastrointestinal bleeding with adjusted Cox proportional hazards models in terms of hazard ratios and 95% confidence intervals for gastrointestinal bleeding with the drug exposure continuously updating—that is, as time varying exposure allocated according to treatment regimen. We considered patients to be at risk only when exposed to the drug (during active treatment). Each patient could have multiple treatment groups throughout follow-up. We calculated risk time (person years) only for the active treatment period. The timescale in the Cox model was days passed since inclusion. We adjusted all models for age, sex, year of index hospital admission, concomitant drugs, comorbidity, and percutaneous coronary intervention status. We did additional analyses to assess any association between PPI use and individual antithrombotic regimens and NSAIDs.
We did nine sensitivity analyses. (1) To take account of any effect of over the counter NSAID use, we ended follow-up at 2001. (2) To take account of dabigatran or ticagrelor use (released in Denmark, August 2011), we ended follow-up in December 2010. (3) We examined cardiovascular death as a solo endpoint. (4) We stratified the cohort at 65 years to take account of guidelines recommending PPIs for people over 65 taking antithrombotic treatment. (5) We controlled for the variables included in the HAS-BLED score. (6) We did an analysis including all covariates as time dependent. (7) We stratified the population in two groups: high (previous bleeding) and low (no previous bleeding) risk of gastrointestinal bleeding. (8) Although the indication for NSAID use was not systematically available, we were able to do an analysis of patients with rheumatoid arthritis. (9) We examined the effect of duration of NSAID treatment (0–14 days and >14 days).
We confirmed the validity of the proportional hazard assumption, linearity of continuous variables, and lack of interaction and chose a significance level of 0.05. We used SAS 9.2 and Stata 11.0 for all statistical calculations.
Patient Involvement
No patients were involved in setting the research question or the outcome measures; nor were they involved in the design and implementation of the study. There are no plans to involve patients in dissemination.
